diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index 05205e6..c33415d 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "idstack",
-  "version": "3.5.1.0",
+  "version": "3.6.0.0",
   "description": "Evidence-based instructional design pipeline. 11 skills backed by peer-reviewed research across 11 domains. Analyze, design, build, review, and export courses with every recommendation citing its evidence tier.",
   "author": {
     "name": "Philippos Savvides",
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 8df0b6a..9434855 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -26,6 +26,13 @@ jobs:
             echo "plugin.json version does not match VERSION ($VER)" >&2; exit 1; }
           grep -qF "## v$VER" CHANGELOG.md || {
             echo "CHANGELOG.md has no entry for v$VER" >&2; exit 1; }
+      # test/test-extension.sh imports the shipped extension ES modules. extension/ has no
+      # package.json, so Node must detect ESM syntax in .js files itself (default only from
+      # 20.19 / 22.7). Pinned so the runner image's Node cannot change under the suite.
+      # Required here too: smoke-test.sh runs test-extension.sh.
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '22'
       - name: Run test suite
         run: ./test/smoke-test.sh
       - name: Create GitHub Release
diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 129c9c3..f4a884a 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -27,14 +27,13 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
-      # smoke-test skips the Chrome-extension and landing-page suites when node is
-      # absent. Declaring it here keeps that skip from silently hiding a failure on a
-      # runner image that stops preinstalling node.
+      # Node 22 is pinned for two suites. test/test-extension.sh imports the shipped extension
+      # ES modules; extension/ has no package.json, so Node must detect ESM syntax in .js files
+      # itself (the default only from 20.19 / 22.7). test-rendered-landing.js drives Chrome over
+      # the global WebSocket, which node exposes unflagged only from 22.4. Declaring node here
+      # also keeps smoke-test's node-absent SKIP from silently hiding a failure.
       - uses: actions/setup-node@v4
         with:
-          # 22, not 20: test-rendered-landing.js drives Chrome over the global WebSocket,
-          # which node exposes unflagged only from 22.4. On 20 it fails with
-          # "WebSocket is not defined" and the rendered suite never runs.
           node-version: '22'
       - name: Smoke test
         run: ./test/smoke-test.sh
@@ -54,6 +53,8 @@ jobs:
         run: ./test/test-preamble-python.sh
       - name: Setup behavioral tests
         run: ./test/test-setup.sh
+      - name: Chrome extension unit tests
+        run: ./test/test-extension.sh
 
   # Confirms the guards above actually fail when their bugs return. Slower
   # (copies the repo per mutation), so it runs once rather than per matrix leg.
@@ -69,10 +70,10 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: '3.9'
-      # Mutations 23-32 are guarded only by the node landing-page suites. Without node,
-      # smoke-test skips them, those mutations report NOT-GUARDED and this job fails. node is
-      # a hard requirement here, not a hedge against a silent skip. 22 for the same WebSocket
-      # reason as the test job; mutation 32 additionally needs Chrome.
+      # Node 22 for the same two reasons as the test job, and a hard requirement here: fresh()
+      # copies extension/ and smoke runs test-extension.sh in every case, and the landing-page
+      # mutations are guarded only by the node suites, so without node they report NOT-GUARDED.
+      # The rendered-landing mutation additionally needs Chrome.
       - uses: actions/setup-node@v4
         with:
           node-version: '22'
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 536dfff..49cb773 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,70 @@
 # Changelog
 
+## v3.6.0.0 (2026-09-24)
+
+Adds an optional Consensus evidence engine to the skills and the Chrome extension, and hardens the extension after a code review of v3.5.0.0. The extension no longer asks for access to every website, refuses Canvas pages that show student records, reads every page of a course's assignments, and its privacy notes now say what is sent to Google and to Consensus. The extension is now version 1.1.0 and needs Chrome 116 or later.
+
+To update: `cd` into your idstack clone, then `git pull && ./setup`. The Chrome Web Store updates the extension on its own once 1.1.0 is published there.
+
+### Added — Consensus evidence engine (bring your own key)
+
+- **`bin/idstack-consensus`** looks up a claim in the Consensus research database and caches every answer under `~/.idstack/cache/consensus/`, so a claim is looked up once. Its subcommands are `status`, `configure`, `query`, `verify`, and `sync`. It reads your own Consensus API key from `CONSENSUS_API_KEY`, `~/.idstack/profile.yaml`, or the project manifest. Without a key it works from `evidence/references.md` alone.
+- **An evidence check before findings are written.** `/idstack:course-quality-review`, `/idstack:assessment-design`, and `/idstack:red-team` now pass their findings through `idstack-consensus verify` before writing the report or the manifest. It corrects four known neuromyths (learning styles, left- and right-brain learning, "10% of the brain", and Dale's cone retention percentages), checks each tier against `evidence/references.md`, and attaches literature matches when a key is set.
+- **Every skill's preamble reports whether a Consensus key is configured**, and `bin/idstack-doctor` checks the CLI and its cache.
+- **The Chrome extension takes an optional Consensus API key in Settings.** With a key, each finding's claim is looked up on Consensus after the audit, and a matched finding shows a Consensus badge and a link to the top paper. The neuromyth corrections apply with or without a key. PRIVACY.md says what is sent and where results are kept.
+
+### Fixed — the extension could read every website
+
+- **v3.5.0.0 injected its page reader into every site you visited.** Chrome's install prompt therefore said it could read your data on all websites, and the same match let the extension's background worker make logged-in requests to any site and read the replies. The reader is now injected only into the tab you are looking at, after you click the idstack toolbar icon. Canvas at `instructure.com` needs no click. A Canvas site on its own domain is requested for that site only, when you first click **Audit Entire Course** there, and Chrome asks you before granting it. The toolbar icon now opens the panel and no longer closes it.
+- **The background worker took course-audit requests from any part of the extension, including the page reader it injected into every site, and fetched whatever course address it was given.** Web pages themselves could not message it. It now ignores messages that don't come from the extension's own pages. The course audit accepts only a bare `https` site address and a numeric course id before it requests anything with your Canvas login.
+
+### Fixed — student-record pages, Modules pages, and Google Docs
+
+- **Canvas pages that show student records were read like any other page.** Gradebook and SpeedGrader, grades, People (in a course or the whole account), groups, discussions, submissions, and Inbox are now refused, and the panel says why instead of auditing them. Other Canvas pages no longer fall back to reading the whole page body, and rubric text is now included. On a Canvas site outside `instructure.com`, only numeric `/courses/<n>` pages are read as Canvas; its other pages, such as the dashboard, are read like any web page.
+- **A Modules page audited only its first item.** Every module item title is now read, including a course home shown in Modules view.
+- **Google Docs were read from the editor's rendered page.** The extension now downloads the document's plain-text export with your Google login, waits at most 10 seconds, accepts only plain text, and says so when a document can't be exported. Published (`/d/e/`) documents are not fetched.
+
+### Fixed — the course audit read at most 50 assignments and hid failures
+
+- **Courses with more than 50 assignments were silently cut short.** The crawler now follows Canvas's pagination to the end, up to 10 pages (500 assignments), and says so when a course has more. A failed request or an unexpected response now stops the audit with an error. It used to become an empty assignment list, so the audit ran as if the course had no assessments.
+- **The course prompt could cut an assignment off partway through.** It keeps its 20,000-character budget but fills it with whole assignments, and its header now states how many of the course's items are shown in full.
+
+### Fixed — empty pages, malformed answers, slow requests, and the API key in the URL
+
+- **An empty or unreadable page was still audited.** Pages with fewer than 10 words are refused before anything is sent, with the reason when the extension knows it (for example, a tab it cannot read). A course with no syllabus text and no assignments is refused too. Both checks also apply in demo mode.
+- **A malformed answer from the model was saved and could crash the results view.** An answer must contain a list of findings before it is saved or shown.
+- **A slow model request had no limit of its own.** It now stops after 25 seconds with a message that says so.
+- **Your API key travelled in the request URL.** It is now sent in a request header.
+
+### Fixed — the side panel labelled, kept, or retried the wrong result
+
+- **A result could carry the wrong page's name.** A page audit is now labelled with the page that was sent, not whichever tab is active when the answer arrives, and a course audit is labelled with the course. A slow tab read can no longer overwrite a newer one.
+- **After an error, the previous result stayed actionable.** The error card now clears it and hides **Add to Dossier** and **Export .md** until the next successful audit. **Retry** re-runs the audit you started, page or course, instead of always running a page audit. If the tab now shows a different course, or no course, Retry goes back to the ready screen instead of auditing it.
+- **A thumbs-up or thumbs-down vote also removed Audit Another Page.** Now only the vote buttons go.
+- **The copy buttons did not report a refused clipboard.** One said "Copied!" anyway and the other failed silently. Both now say "Copy failed".
+- **The results view crashed on findings that were not a list, on empty entries, or on a tier that was not text.** It now handles all three.
+
+### Fixed — evidence labels disagreed with `evidence/references.md`
+
+- **The extension's tier scale put randomized trials in T2 and called T3 observational.** Its labels now match idstack's scale, T1 meta-analyses and RCTs through T5 expert opinion, and both prompts give the model that scale.
+- **Demo findings and the course prompt cited studies idstack does not hold, or held them at the wrong tier.** Biggs (1996) is now `[Alignment-1] [T5]`, not `[Alignment-3]` at T2. The non-existent `[Cognitive-2]` is replaced by `[CogLoad-1]` and `[CogLoad-6]`, which exist. Carpenter (2022), Sweller (2011), and Wood (1976), none of which are in `evidence/references.md`, are replaced by studies that are. An unused list of evidence domains that cited unreferenced studies is removed.
+- **The page prompt asked for a `suggestion` severity, which idstack does not use.** It now asks for `critical`, `warning`, or `info`.
+
+### Fixed — the Chrome extension's privacy promises did not match what it does
+
+- **The Settings panel promised that no student data is "ever collected, retained, or used for model training."** None of that held. On a Canvas page the extractor did not recognize, such as Gradebook, People, Inbox, or a discussion, it read the whole page. The panel also points you to a free Google AI Studio key, whose terms let Google use submitted content to improve its products and let human reviewers read it. The extractor now refuses those pages (see above), and the panel says what is sent, to whom, and what Google's free tier allows.
+- **"Leave blank to use the built-in free demo tier" described a service that does not exist.** Without a key, every audit returns the same sample findings, whatever the page says. The panel, README, and PRIVACY.md now say so.
+- **PRIVACY.md said the API key is stored with `chrome.storage.local`.** The code uses `chrome.storage.sync`, which Chrome copies to your Google account when sync is on. The policy now says that. It also discloses that the last 20 audit results are kept in `chrome.storage.local` and that the side panel loads its fonts from Google Fonts. It no longer claims that the extension never touches rosters or grades, or that its design is FERPA-compliant. The README's "Is my data safe?" answer now points to it for the extension.
+- **Correction to the v3.5.0.0 notes.** They said the course crawler audits "syllabi, modules, assignments, discussions, quizzes". It requests the syllabus and assignments only. README repeated the claim, along with a course "dashboard" and a JSON export that do not exist, and is fixed. The v3.5.0.0 entry is left as released.
+
+### For contributors
+
+- **The mutation suite refuses a mutation that changes nothing.** Two cases whose anchors had gone stale kept reporting GUARDED because nothing checked that the mutation had edited anything. The suite now copies the repo once into a snapshot that its baseline smoke run vouches for, and stops on any mutation that leaves the copy unchanged. It grows from 36 cases to 108.
+- **The retired-CLI sweep now covers `docs/superpowers/`.** `--exclude-dir=superpowers` matched that name at any depth and exempted every committed spec and plan.
+- **The extension tests run the shipped code.** They used to load hand-copied `.cjs` twins, which are gone, and `test-extension.sh` fails if one comes back. `test/extension-harness.mjs` stubs the Chrome APIs and a small DOM, so the tests import the shipped modules and now also run the background worker, storage, and side panel. `test/test-evidence-labels.mjs` checks every citation and tier against `evidence/references.md`. `test/test-disclosures.mjs` checks the side panel, PRIVACY.md, and README against the shipped code.
+- **The release workflow pins Node 22**, as the test and mutation jobs already did, and `test-extension.sh` stops with a clear message on a Node too old to import the modules (it needs 20.19+ or 22.7+). The extension suite also has its own CI step, so its full failure output shows.
+- **`bin/package-extension.sh` names the zip from `extension/manifest.json`** and leaves the dev-only `icons/generate-icons.js` out of it. The side panel's version badge is tested against the manifest.
+
 ## v3.5.1.0 (2026-08-29)
 
 Fixes three defects on idstack.org that only show up on a phone, and repairs the test harness that
diff --git a/CLAUDE.md b/CLAUDE.md
index 1556d04..51d089e 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -27,7 +27,7 @@ bin/package-extension.sh                   # Package Chrome extension into build
 Tests (run in CI on every push and PR — see `.github/workflows/test.yml`):
 
 ```bash
-./test/smoke-test.sh              # 362 assertions: install, SKILL.md freshness, frontmatter, version agreement,
+./test/smoke-test.sh              # 365 assertions: install, SKILL.md freshness, frontmatter, version agreement,
                                   # canonical section names, /idstack: namespacing, resolve-snippet lockstep, bash -n,
                                   # Claude-Code-only invariant (no dist/, no AGENTS.md, no retired-CLI references)
 ./test/integration-test.sh        # 51 behavioral tests across the bin/ scripts; also proves the suite
@@ -39,7 +39,8 @@ Tests (run in CI on every push and PR — see `.github/workflows/test.yml`):
 ./test/test-version-classifier.sh # bin/lib/version-classify.sh unit tests
 ./test/test-plugin-status.sh      # bin/lib/plugin-status.sh unit tests
 ./test/test-preamble-python.sh    # Runs the preamble's embedded python on 3.9 and 3.12
-./test/test-extension.sh          # Chrome Extension packaging and static analysis tests
+./test/test-extension.sh          # Chrome extension unit tests; import the shipped extension/ modules via
+                                  # test/extension-harness.mjs (Node >= 20.19 or 22.7; never add .cjs twins)
 ./test/test-responsive-landing.js # Responsive/mobile invariants for docs/index.html, as CSS text (node; via smoke-test)
 ./test/test-rendered-landing.js   # Renders docs/index.html in headless Chrome and asserts the outcome:
                                   # no sideways scroll, 44px touch targets, column counts (node + Chrome)
@@ -48,9 +49,9 @@ python3 test/check-doc-accuracy.py .   # Verifies documentation accuracy across
 ./test/mutation-test.sh           # Reintroduces each fixed defect and asserts its guarding test fails
 ```
 
-`test/test-helper.sh` is not a suite — it is sourced by every **bash** suite that runs assertions itself and owns the shared `PASS`/`FAIL`/`TOTAL` counters and the `check()` assertion (`test-extension.sh` is the exception: it delegates to node and asserts nothing of its own). Do not add a local counter block to a bash suite; smoke-test fails on one, and a mutation proves that guard works. The node suites (`test-extension.sh`'s eight unit tests, `test/test-responsive-landing.js`, and `test/test-rendered-landing.js`) cannot source it; they accumulate their own problems and report a count, which is the same contract in another language.
+`test/test-helper.sh` is not a suite — it is sourced by every **bash** suite that runs assertions itself and owns the shared `PASS`/`FAIL`/`TOTAL` counters and the `check()` assertion (`test-extension.sh` is the exception: it delegates to node). Do not add a local counter block to a bash suite; smoke-test fails on one, and a mutation proves that guard works. The node suites (`test-extension.sh`'s sixteen unit tests, `test/test-responsive-landing.js`, and `test/test-rendered-landing.js`) cannot source it; they accumulate their own problems and report a count, which is the same contract in another language.
 
-CI matrix: ubuntu (Python 3.9 + 3.12) and macOS (3.12). 3.9 is the leg that catches modern-only Python syntax reaching the preamble's embedded scripts — it is what macOS ships. `mutation-test.sh` runs once, pinned to 3.9.
+CI matrix: ubuntu (Python 3.9 + 3.12) and macOS (3.12). 3.9 is the leg that catches modern-only Python syntax reaching the preamble's embedded scripts — it is what macOS ships. `mutation-test.sh` runs once, pinned to 3.9. The test, mutation and release jobs pin Node 22 for the extension and rendered-landing tests.
 
 No build step for users. No dependencies beyond bash (python3 recommended for full features). Skills are plain Markdown files.
 
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 2a2b79c..7d50b6e 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -140,7 +140,7 @@ Every suite below runs in CI (`.github/workflows/test.yml`) on every push and pu
 | `test/test-version-classifier.sh` | `bin/lib/version-classify.sh` |
 | `test/test-plugin-status.sh` | `bin/lib/plugin-status.sh` |
 | `test/test-preamble-python.sh` | The preamble's embedded Python, on 3.9 and 3.12 |
-| `test/test-extension.sh` | Chrome extension packaging and static analysis, plus eight node unit suites under `test/test-*.js` |
+| `test/test-extension.sh` | The Chrome extension. Imports the shipped `extension/` modules through `test/extension-harness.mjs` (Node 20.19+ or 22.7+) and fails if a hand-copied `.cjs` twin comes back |
 | `test/test-rendered-landing.js` | Renders `docs/index.html` in headless Chrome across 11 widths and asserts the rendered outcome: no horizontal scroll, 44px touch targets, breakpoint column counts, sticky nav. Catches regressions the text suite cannot see (selector lists, `@container`, nested or print-only media queries, a `<style>` inside an HTML comment). Skips loudly with no browser |
 | `test/test-responsive-landing.js` | Responsive and mobile-ergonomics invariants for `docs/index.html` — fluid tokens, notch-safe gutters, breakpoint-scoped rules, touch targets. Runs on node, via `smoke-test.sh` |
 | `python3 test/check-evidence-cards.py .` | Verifies landing page evidence card study counts and tier ranges against `evidence/references.md` |
diff --git a/PRIVACY.md b/PRIVACY.md
index f88e5ed..c494262 100644
--- a/PRIVACY.md
+++ b/PRIVACY.md
@@ -1,10 +1,10 @@
 # Privacy Policy
 
-**Last updated:** April 21, 2026
+**Last updated:** September 24, 2026
 
 ## What idstack collects
 
-Nothing. idstack runs entirely on your local machine.
+Nothing. idstack has no servers of its own and receives none of your data. The Claude Code plugin runs on your machine and the Chrome extension runs in your browser. The extension sends page text to Google only when you run an audit with your own Google API key saved, and both send the text of findings to Consensus only when you have set up your own Consensus key (see below).
 
 ## Where your data lives
 
@@ -20,7 +20,7 @@ Designer profile and cross-project learnings are stored locally:
 - `~/.idstack/profile.yaml` — your experience level preference
 - `~/.idstack/global/learnings.jsonl` — learnings promoted across projects
 
-idstack adds no analytics, no tracking, and no telemetry. Two things do reach the network, both only when you ask for them, and both are listed under Third-party services below: the Canvas API calls made by `/idstack:course-import` and `/idstack:course-export`, and an update check against this repository on GitHub.
+idstack adds no analytics, no tracking, and no telemetry. In the Claude Code plugin, three things can reach the network, all listed under Third-party services below: the Canvas API calls made by `/idstack:course-import` and `/idstack:course-export`, an update check against this repository on GitHub, and Consensus lookups once you set up a Consensus key.
 
 ## Third-party services
 
@@ -33,16 +33,23 @@ idstack runs inside Claude Code, which is operated by Anthropic. Your conversati
 
 Both run only when you invoke that skill and confirm the target course. Your token is used for those API calls and is neither stored by idstack nor sent anywhere else. The receiving Canvas instance is your institution's, not ours.
 
+**Consensus API.** Only if you set up your own Consensus API key: `CONSENSUS_API_KEY` in your environment, `consensus_api_key` in `~/.idstack/profile.yaml` (where `bin/idstack-consensus configure` saves it), or `preferences.consensus_api_key` in `.idstack/project.json`. With a key, `bin/idstack-consensus` sends the text of each claim it checks to Consensus (`api.consensus.app`) under your key. Those claims are the findings from `/idstack:course-quality-review`, `/idstack:assessment-design`, and `/idstack:red-team`, which describe your course and can quote it, or the research domain you name to `idstack-consensus sync`. Answers are cached in `~/.idstack/cache/consensus/`, and a claim already in the cache is not sent again. A key kept in `.idstack/project.json` travels with that file, so if you share the project folder, consider the environment variable or your profile instead. Without a key, nothing is sent to Consensus.
+
 **Update check.** On skill startup idstack runs `git fetch` against this repository to see whether a newer version exists, at most once an hour. That is a request to GitHub carrying nothing but the fetch itself; it never uploads your course data. It only runs for git installs, and removing the repo's `.git` directory disables it.
 
 ## idstack Chrome Extension
 
-The idstack Chrome Extension is designed with a strict privacy-first and FERPA-compliant architecture:
-
-- **Curriculum-Only Processing:** The extension only reads public or instructor-accessible course materials (syllabi, module structures, assignment guidelines, and rubrics). It never accesses student rosters, student submissions, student grades, or any Personally Identifiable Information (PII).
-- **Client-Side Storage:** Your optional Google AI Studio API key and saved course audit dossiers are stored locally on your device via `chrome.storage.local`. No audit history or credentials are ever sent to idstack servers.
-- **Direct AI Inference:** If you provide your own Google AI Studio API key, requests are sent directly from your browser to Google's API (`generativelanguage.googleapis.com`). Zero data is routed through intermediary proxy servers.
-- **No Tracking:** The extension contains zero analytics, tracking scripts, or telemetry.
+The extension is built to audit course materials, not student records. Here is what it reads, where that text goes, and what it keeps.
+
+- **What it reads.** The extension reads a tab only when it has access: after you click the idstack toolbar icon on that tab, on Canvas pages at `instructure.com`, or on a Canvas site you allowed when Chrome asked. While the side panel is open it reads the title and main text of such a tab; that text stays in the panel until you click **Audit Page with Evidence**. On a Google Doc it downloads the document's plain text using your Google login. **Audit Entire Course** requests the course syllabus and assignments from your Canvas instance's API using your existing Canvas login. The extension will not read Canvas grades, Gradebook, SpeedGrader, People, groups, discussions, submissions, or Inbox pages. Other pages can still contain student names or work, so audit course materials rather than student work.
+- **Demo mode.** With no Google API key saved, the page's text is not sent anywhere: every audit returns the same sample findings, whatever the page says. If a Consensus key is saved, those sample findings are still checked against Consensus.
+- **Where the text goes with a key.** With your own Google AI Studio API key saved, the page title and text (or the syllabus and assignments) are sent directly from your browser to Google's Generative Language API (`generativelanguage.googleapis.com`) under your key and your agreement with Google. No idstack server is involved.
+- **Google's free tier.** If your key's Google Cloud project has no active Cloud Billing account, Google's terms let Google use what you submit, and the responses, to improve its products, and human reviewers may read them. The same terms ask you not to submit personal information. On billed projects, and for users in the European Economic Area, Switzerland, and the United Kingdom, Google says it does not use prompts or responses to improve its products. See [Google's terms for this API](https://ai.google.dev/gemini-api/terms). <!-- IDSTACK_CLI_LEAK_ALLOW -->
+- **Consensus, with a Consensus key.** If you also save a Consensus key in Settings, then after each audit the claim of every finding (its evidence text, or failing that its observation, recommendation, or citation) is sent from your browser to Consensus (`api.consensus.app`) under your key. Findings describe the audited page and can quote it. Answers are cached in `chrome.storage.local`, and a claim already in the cache is not sent again. Without a Consensus key, nothing goes to Consensus.
+- **API keys.** Your Google API key and your Consensus key are saved in `chrome.storage.sync`. If Chrome sync is on, Chrome copies them to your Google account and to other browsers where you are signed in. The extension sends each key only to its own service, with each audit. To remove one, clear its field in Settings and save.
+- **Audit history and dossier.** The last 20 audit results (page address, title, page type, and the full audit result) are kept in `chrome.storage.local`, as is the Course Dossier you build. Both stay on this device. **Clear Dossier** empties the dossier; removing the extension deletes both.
+- **Fonts.** Opening the side panel loads its fonts from Google Fonts (`fonts.googleapis.com`, `fonts.gstatic.com`), so Google receives your IP address and browser details; no page or course content is sent.
+- **No tracking.** The extension contains no analytics, tracking scripts, or telemetry.
 
 ## Questions
 
diff --git a/README.md b/README.md
index 8929248..965126b 100644
--- a/README.md
+++ b/README.md
@@ -1,6 +1,6 @@
 # idstack
 
-> **Status: beta (v3.5.1.0).** Skills work end-to-end and ship behind a versioned plugin, but expect rough edges and breaking changes between minor versions while we shake out the rest of the workflow with real instructional designers. See [CHANGELOG.md](CHANGELOG.md) for what's new in the current release. [Tell us what's broken or missing.](https://forms.gle/6LDgDD1M6WWyYvME8)
+> **Status: beta (v3.6.0.0).** Skills work end-to-end and ship behind a versioned plugin, but expect rough edges and breaking changes between minor versions while we shake out the rest of the workflow with real instructional designers. See [CHANGELOG.md](CHANGELOG.md) for what's new in the current release. [Tell us what's broken or missing.](https://forms.gle/6LDgDD1M6WWyYvME8)
 
 Decades of research say elaborated feedback improves learning. Cognitive load theory has been replicated for 30 years. Constructive alignment measurably raises grades. The evidence is strong, but most course design tools don't make it easy to apply.
 
@@ -188,6 +188,8 @@ If you or your instructional designers don't use the terminal, idstack is availa
 3. Navigate to any Canvas assignment, syllabus, or Google Doc and click the idstack icon to open the Side Panel.
 4. Click **"Audit Page with Evidence"** for instant Bloom's classification, constructive alignment reviews, and 1-click improved rubrics.
 
+If you switch tabs while the panel is open, click the idstack icon again so it can read the new tab. This isn't needed on Canvas pages at `instructure.com` or on a Canvas site you've already allowed.
+
 #### Loading Unpacked (Local Development)
 
 If developing or modifying the extension locally:
@@ -198,24 +200,26 @@ If developing or modifying the extension locally:
 #### Full Course Audit (Canvas LMS)
 
 When viewing any Canvas course homepage or modules list (`/courses/:id`), idstack automatically detects the course environment and presents an **"Audit Entire Course"** button:
-- **Zero Developer Tokens Required:** Crawls published syllabus items, modules, assignments, discussions, and quizzes in the background using your active browser session. No Canvas API keys, LMS admin setup, or command line required.
-- **Course Quality Matrix:** Synthesizes findings across all modules into an interactive course health dashboard, highlighting cognitive process distribution across Bloom's levels, constructive alignment gaps, and prioritized action items.
-- **Export Ready:** Download the full synthesized course audit as JSON or Markdown to share with instructional design teams and faculty stakeholders.
+- **Zero Developer Tokens Required:** Reads the course syllabus and assignment descriptions through Canvas's API using your active browser session. No Canvas API keys, LMS admin setup, or command line required.
+- **Course-level findings:** Reviews the syllabus and assignments together for alignment gaps, cognitive demand across Bloom's levels, and prioritized fixes, shown in the same results view as a page audit.
+- **Export Ready:** Add the course audit to your Course Dossier or export it as Markdown to share with instructional design teams and faculty stakeholders.
+
+On Canvas hosted outside `instructure.com` (for example, a university's own domain), Chrome asks once, for that site only, when you first click **"Audit Entire Course"**.
 
 #### Multi-Page Course Dossier & Markdown Export
 
 Audit multiple pages across a course or syllabus and accumulate findings into a structured institutional deliverable:
-- **Incremental Dossier Accumulation:** Click **"Add to Dossier"** on any page audit (assignments, quizzes, syllabi, Google Docs). The header dossier badge tracks your collected course components across browsing sessions.
+- **Incremental Dossier Accumulation:** Click **"Add to Dossier"** on any page audit (assignments, pages, syllabi, rubrics, Google Docs). The header dossier badge tracks your collected course components across browsing sessions.
 - **1-Click Compiled Markdown Export (`.md`):** Open the Dossier drawer to download a synthesized multi-page Markdown report (`idstack-course-dossier-<title>.md`) containing executive summaries, cognitive demand tables, itemized empirical citations (T1–T5), and improved rubric drafts.
 - **Clipboard Ready:** Copy the compiled Markdown report directly to your clipboard for pasting into LMS course notes, Google Docs, Notion, or faculty review tickets.
 
 #### Free Google AI Studio API Key Setup
 
-idstack includes a rich built-in simulation fallback mode for instant demonstration without API keys. To connect live LLM inference:
+Without a key, the extension runs in demo mode: every audit returns the same sample findings, whatever the page says. To get real audits:
 1. Obtain a free API key from [Google AI Studio](https://aistudio.google.com/app/apikey).
 2. In the idstack Side Panel, click the **Settings** (gear) icon in the top header.
 3. Paste your API key into the input field and click **Save Settings**.
-4. All prompts execute client-side directly against Google's API (`gemini-2.5-flash`). Zero student PII or course content is stored or transmitted to external intermediary servers. <!-- IDSTACK_CLI_LEAK_ALLOW -->
+4. Each audit sends the page's title and text (or, for a full-course audit, the syllabus and assignments) straight from your browser to Google's API; no idstack server is involved. On Google's free tier, Google may use that content to improve its products, so read [PRIVACY.md](PRIVACY.md) before auditing anything that names students.
 
 #### Packaging for Chrome Web Store
 
@@ -223,7 +227,7 @@ To create a distribution bundle:
 ```bash
 ./bin/package-extension.sh
 ```
-Produces `build/idstack-chrome-extension-v1.0.0.zip` ready for submission to the Chrome Web Store Developer Console.
+Produces `build/idstack-chrome-extension-v<version>.zip` (version from `extension/manifest.json`), ready for submission to the Chrome Web Store Developer Console.
 
 
 ## Your design team
@@ -433,7 +437,7 @@ Run `bin/idstack-status` from your project directory for the dashboard (quality
 python3 is recommended but not required. With python3, you get quality score trends, dimension analysis, search filtering, and safe JSON serialization. Without it, basic timeline logging and learnings still work via bash fallback, but score trends and search filtering are unavailable. Most systems have python3 pre-installed.
 
 **Is my data safe?**
-idstack runs locally on your machine. Course data stays in your project folder (`.idstack/project.json`). Session history (`.idstack/timeline.jsonl`) and learnings (`.idstack/learnings.jsonl`) are also local. Course data reaches external servers only when you explicitly invoke integration skills (such as Canvas API calls via `/idstack:course-import` or `/idstack:course-export`) or perform hourly update checks via `git fetch`. Claude Code's own privacy policy applies to the AI conversation.
+idstack runs locally on your machine. Course data stays in your project folder (`.idstack/project.json`). Session history (`.idstack/timeline.jsonl`) and learnings (`.idstack/learnings.jsonl`) are also local. In the Claude Code plugin, course data reaches external servers only when you explicitly invoke integration skills (such as Canvas API calls via `/idstack:course-import` or `/idstack:course-export`) or perform hourly update checks via `git fetch`, and findings are sent to Consensus only if you set up your own Consensus key. The Chrome extension sends page text to Google's API when you run an audit with your own API key saved, and finding text to Consensus when you also save a Consensus key; see [PRIVACY.md](PRIVACY.md). Claude Code's own privacy policy applies to the AI conversation.
 
 **Can I edit the project manifest directly?**
 You can, but let the skills manage it. If you do edit it, keep the JSON valid.
diff --git a/ROADMAP.md b/ROADMAP.md
index 5866dc9..7cf70fc 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -4,6 +4,11 @@ What's coming next for idstack. Priorities are shaped by user feedback. [Tell us
 
 ## Just shipped
 
+### Consensus evidence checks & Chrome extension hardening (v3.6.0.0)
+- **An optional Consensus evidence engine, on your own API key.** `bin/idstack-consensus` looks up and caches claims. `/idstack:course-quality-review`, `/idstack:assessment-design`, and `/idstack:red-team` check their findings with it before writing them, and correct four known neuromyths with or without a key.
+- **The Chrome extension (1.1.0) reads only the tab you point it at.** It no longer asks for access to every website, refuses Canvas pages that show student records, and reads every page of a course's assignments.
+- **Its privacy notes match the code.** The side panel and PRIVACY.md say what is sent to Google and to Consensus, and where results are kept.
+
 ### Mobile fixes for idstack.org & test-harness repair (v3.5.1.0)
 - **idstack.org is usable on a phone.** Every content section sat under the display cutout, buttons and inputs fell below the 44px touch floor at every width above 480px, and a viewport clip was hiding horizontal overflow instead of preventing it. Measured clean at 11 widths from 320px up. No skill behavior changed.
 - **The mutation suite could not fail.** `test/mutation-test.sh` never copied `extension/` into its throwaway repos, so `test/test-extension.sh` broke before any mutation was applied and all 36 mutations reported GUARDED whether their guard worked or not. A null-mutation control now aborts the run when an unmutated copy is already red.
diff --git a/VERSION b/VERSION
index a11b5ad..68cd19e 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-3.5.1.0
+3.6.0.0
diff --git a/bin/package-extension.sh b/bin/package-extension.sh
index 15daf6d..408ebf2 100755
--- a/bin/package-extension.sh
+++ b/bin/package-extension.sh
@@ -7,7 +7,8 @@ SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
 EXT_DIR="$REPO_ROOT/extension"
 BUILD_DIR="$REPO_ROOT/build"
-ZIP_NAME="idstack-chrome-extension-v1.0.0.zip"
+# The zip is named for the version in manifest.json, so a release bump cannot ship a stale name.
+ZIP_NAME="idstack-chrome-extension-v$(node -p 'require(process.argv[1]).version' "$EXT_DIR/manifest.json").zip"
 ZIP_PATH="$BUILD_DIR/$ZIP_NAME"
 
 echo "==> Verifying extension test suite..."
@@ -19,7 +20,8 @@ rm -f "$ZIP_PATH"
 
 echo "==> Packaging extension from $EXT_DIR..."
 # Zip contents directly so manifest.json is at root of archive
-(cd "$EXT_DIR" && zip -r "$ZIP_PATH" . -x "*.DS_Store" "*__MACOSX*" "*.git*")
+# icons/generate-icons.js is a dev-time script that drew the PNGs; it is not part of the extension.
+(cd "$EXT_DIR" && zip -r "$ZIP_PATH" . -x "*.DS_Store" "*__MACOSX*" "*.git*" "icons/generate-icons.js")
 
 echo ""
 echo "✅ Extension packaged successfully!"
diff --git a/docs/index.html b/docs/index.html
index b8ff657..78c9df0 100644
--- a/docs/index.html
+++ b/docs/index.html
@@ -42,7 +42,7 @@
     "url": "https://idstack.org",
     "applicationCategory": "EducationalApplication",
     "operatingSystem": "macOS, Linux, Windows (WSL or Git Bash)",
-    "softwareVersion": "3.5.1.0",
+    "softwareVersion": "3.6.0.0",
     "datePublished": "2026-04-20",
     "offers": {
       "@type": "Offer",
@@ -1030,7 +1030,7 @@
           <span class="dot" aria-hidden="true"></span>
           <span>MIT</span>
           <span class="dot" aria-hidden="true"></span>
-          <span>v3.5.1.0 · Claude Code &amp; Chrome Extension</span>
+          <span>v3.6.0.0 · Claude Code &amp; Chrome Extension</span>
         </p>
         <h1 id="hero-title">Evidence-based instructional design with peer-reviewed citations.</h1>
         <p class="lede">108 peer-reviewed studies across 11 research domains. Every recommendation includes its evidence tier: <span class="tier tier-1">T1</span> meta-analyses and RCTs through <span class="tier tier-5">T5</span> expert opinion. Runs in Claude Code and as a native Chrome Extension.</p>
@@ -1318,6 +1318,13 @@ <h3>Claude Code Plugin</h3>
         <h2 id="whatsnew-title">Recent Updates &amp; Release Notes</h2>
       </header>
 
+      <article class="release-card">
+        <p class="release-meta">September 24, 2026 · v3.6.0.0</p>
+        <h3>Consensus evidence checks &amp; a hardened Chrome extension.</h3>
+        <p>An optional Consensus evidence engine runs on your own API key. <code>/idstack:course-quality-review</code>, <code>/idstack:assessment-design</code>, and <code>/idstack:red-team</code> now check their findings before writing them: four well-known neuromyths are corrected with or without a key, and with one each claim is matched against the research literature. The Chrome extension can run the same check and marks the findings it matches.</p>
+        <p>The Chrome extension, now 1.1.0, no longer requests access to every website, refuses Canvas pages that show student records, reads every page of a course's assignments, and its privacy notes now say what is sent to Google and to Consensus.</p>
+      </article>
+
       <article class="release-card">
         <p class="release-meta">August 16, 2026 · v3.5.0.0 · patched through v3.5.1.0</p>
         <h3>Chrome Side Panel co-pilot &amp; Course Dossier export.</h3>
diff --git a/docs/superpowers/plans/2026-08-15-chrome-extension-experiment.md b/docs/superpowers/plans/2026-08-15-chrome-extension-experiment.md
index 0be01cf..41363c0 100644
--- a/docs/superpowers/plans/2026-08-15-chrome-extension-experiment.md
+++ b/docs/superpowers/plans/2026-08-15-chrome-extension-experiment.md
@@ -4,7 +4,7 @@
 
 **Goal:** Build a frictionless Manifest V3 Chrome Extension that brings idstack's evidence-based course design audits into Canvas LMS, Google Docs, and web browsers via Chrome's native Side Panel.
 
-**Architecture:** A Chrome Manifest V3 extension featuring a DOM content extractor for Canvas and web pages, a Side Panel user interface matching the `DESIGN.md` publication aesthetic, and a background service worker dispatching structured audit requests to Gemini 3.7 Flash with T1–T5 evidence citations.
+**Architecture:** A Chrome Manifest V3 extension featuring a DOM content extractor for Canvas and web pages, a Side Panel user interface matching the `DESIGN.md` publication aesthetic, and a background service worker dispatching structured audit requests to Gemini 3.7 Flash with T1–T5 evidence citations. <!-- IDSTACK_CLI_LEAK_ALLOW -->
 
 **Tech Stack:** Chrome Extensions API (Manifest V3, Side Panel API, Content Scripts, Service Worker), Vanilla JavaScript (ES Modules), Vanilla CSS (matching `DESIGN.md` tokens), Node.js test runner for unit tests.
 
@@ -407,7 +407,7 @@ git commit -m "feat(extension): implement DOM extractor for Canvas and web sylla
 
 ---
 
-### Task 4: Background Service Worker & Gemini API Engine
+### Task 4: Background Service Worker & Gemini API Engine <!-- IDSTACK_CLI_LEAK_ALLOW -->
 
 **Files:**
 - Create: `extension/background/service-worker.js`
@@ -424,9 +424,9 @@ Create `test/test-service-worker.js`:
 const assert = require('assert');
 const { parseAuditResponse } = require('../extension/background/parser-helper.cjs');
 
-const rawGeminiResponse = "```json\n{\n  \"summary\": {\n    \"bloomsLevel\": \"Remember\",\n    \"alignmentScore\": \"Moderate\",\n    \"keyTakeaway\": \"Quiz focuses only on memorization.\"\n  },\n  \"findings\": [],\n  \"improvedDraft\": {\n    \"title\": \"Analysis Prompt\",\n    \"content\": \"Compare and contrast\"\n  }\n}\n```";
+const rawGeminiResponse = "```json\n{\n  \"summary\": {\n    \"bloomsLevel\": \"Remember\",\n    \"alignmentScore\": \"Moderate\",\n    \"keyTakeaway\": \"Quiz focuses only on memorization.\"\n  },\n  \"findings\": [],\n  \"improvedDraft\": {\n    \"title\": \"Analysis Prompt\",\n    \"content\": \"Compare and contrast\"\n  }\n}\n```"; // IDSTACK_CLI_LEAK_ALLOW
 
-const parsed = parseAuditResponse(rawGeminiResponse);
+const parsed = parseAuditResponse(rawGeminiResponse); // IDSTACK_CLI_LEAK_ALLOW
 assert.strictEqual(parsed.summary.bloomsLevel, 'Remember');
 assert.strictEqual(parsed.improvedDraft.title, 'Analysis Prompt');
 
@@ -456,8 +456,8 @@ function cleanJsonResponse(rawText) {
   return JSON.parse(cleaned.trim());
 }
 
-async function callGeminiApi(apiKey, prompt) {
-  const url = `https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=${apiKey}`;
+async function callGeminiApi(apiKey, prompt) { // IDSTACK_CLI_LEAK_ALLOW
+  const url = `https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent?key=${apiKey}`; // IDSTACK_CLI_LEAK_ALLOW
   const response = await fetch(url, {
     method: 'POST',
     headers: { 'Content-Type': 'application/json' },
@@ -492,10 +492,10 @@ chrome.runtime.onMessage.addListener((request, sender, sendResponse) => {
         // Use user's key if provided, or default endpoint
         let auditResult;
         if (settings.apiKey) {
-          auditResult = await callGeminiApi(settings.apiKey, prompt);
+          auditResult = await callGeminiApi(settings.apiKey, prompt); // IDSTACK_CLI_LEAK_ALLOW
         } else {
           // Fallback demo mock or proxy endpoint
-          auditResult = await callGeminiApi('YOUR_DEFAULT_API_KEY_OR_PROXY', prompt);
+          auditResult = await callGeminiApi('YOUR_DEFAULT_API_KEY_OR_PROXY', prompt); // IDSTACK_CLI_LEAK_ALLOW
         }
 
         await saveAuditResult({
@@ -624,7 +624,7 @@ Create `extension/sidepanel/index.html`:
       </div>
       <div class="drawer-body">
         <div class="form-group">
-          <label for="api-key-input">Gemini / Claude API Key (Optional BYOK)</label>
+          <label for="api-key-input">Gemini / Claude API Key (Optional BYOK)</label> <!-- IDSTACK_CLI_LEAK_ALLOW -->
           <input type="password" id="api-key-input" placeholder="AIzaSy... or sk-ant-..." />
           <p class="help-text">Leave blank to use the standard free demo tier.</p>
         </div>
diff --git a/docs/superpowers/plans/2026-08-15-course-crawler.md b/docs/superpowers/plans/2026-08-15-course-crawler.md
index 5063523..323695c 100644
--- a/docs/superpowers/plans/2026-08-15-course-crawler.md
+++ b/docs/superpowers/plans/2026-08-15-course-crawler.md
@@ -13,7 +13,7 @@
 - **Design System:** Strictly follow `DESIGN.md` (Source Serif 4, Public Sans, JetBrains Mono, `#faf8f3` ivory background, `#ffffff` card surfaces, `#1a1815` ink, canonical T1–T5 palette: T1 `#2f7a4a`, T2 `#2864a8`, T3 `#a87726`, T4 `#b35a1f`, T5 `#6b6b6b`).
 - **Zero Build Overhead:** Pure Vanilla JS and CSS without bundlers, webpack, or npm runtime dependencies.
 - **Privacy & FERPA:** Only extract syllabus, assignment descriptions, and rubrics. Never touch student rosters, submissions, or grades.
-- **Free Tier Onboarding:** Provide direct link to Google AI Studio for free Gemini API keys, plus a rich demo fallback mode for instant testing without an API key.
+- **Free Tier Onboarding:** Provide direct link to Google AI Studio for free Gemini API keys, plus a rich demo fallback mode for instant testing without an API key. <!-- IDSTACK_CLI_LEAK_ALLOW -->
 - **T1–T5 Research Grounding:** All course audit prompts ground alignment ratings in `evidence/references.md` (Biggs constructive alignment, Bloom's taxonomy, Hattie feedback, Sweller cognitive load).
 
 ---
diff --git a/docs/superpowers/specs/2026-08-15-chrome-extension-design.md b/docs/superpowers/specs/2026-08-15-chrome-extension-design.md
index 7d8de38..da4a022 100644
--- a/docs/superpowers/specs/2026-08-15-chrome-extension-design.md
+++ b/docs/superpowers/specs/2026-08-15-chrome-extension-design.md
@@ -33,7 +33,7 @@ Building institutional LMS integrations (LTI 1.3, Canvas Developer Keys) introdu
        ▼ (Automatic platform detection: "Canvas Assignment Detected")
 [User clicks "Audit with Evidence"]
        │
-       ▼ (2-3s fast analysis via Gemini 3.7 Flash)
+       ▼ (2-3s fast analysis via Gemini 3.7 Flash) <!-- IDSTACK_CLI_LEAK_ALLOW -->
 [Results Displayed in Side Panel]
  ├── Bloom's Taxonomy Level & Alignment Summary
  ├── Finding Cards with [T1]–[T5] Evidence Badges
@@ -49,7 +49,7 @@ Building institutional LMS integrations (LTI 1.3, Canvas Developer Keys) introdu
    - **Finding Cards:** Categorized by severity (Critical / Warning / Suggestion) with mono-tagged citation chips (e.g. `[Assessment-8] [T1]`).
    - **Actionable Rewrite Box:** Formatted revised rubric criteria or rewritten measurable learning objectives with a **"📋 Copy to Clipboard"** button.
    - **Feedback Mechanism:** Instant 1-click **"Helpful? 👍 / 👎"** toggle to collect experiment signal.
-4. **Settings / Privacy State:** Clean drawer allowing users to view privacy commitments (zero student PII stored) and optionally enter a custom Gemini / Claude API key for BYOK mode.
+4. **Settings / Privacy State:** Clean drawer allowing users to view privacy commitments (zero student PII stored) and optionally enter a custom Gemini / Claude API key for BYOK mode. <!-- IDSTACK_CLI_LEAK_ALLOW -->
 
 ---
 
@@ -88,7 +88,7 @@ The content script is injected on demand or on target domains to extract course
 ### 3.2 Background Worker (`background/service-worker.js`)
 - Receives audit requests from the Side Panel.
 - Formats prompt with extracted DOM text + idstack research criteria.
-- Dispatches request to the configured LLM endpoint (Gemini 3.7 Flash default proxy or local BYOK endpoint).
+- Dispatches request to the configured LLM endpoint (Gemini 3.7 Flash default proxy or local BYOK endpoint). <!-- IDSTACK_CLI_LEAK_ALLOW -->
 - Validates and parses the returned JSON schema and forwards results to the Side Panel.
 
 ---
@@ -139,6 +139,6 @@ For this experiment to be considered successful before expanding into larger fea
 1. **Phase 1: Project Scaffolding & Manifest:** Create `extension/manifest.json`, icon assets, and baseline extension setup.
 2. **Phase 2: Content Extraction Engine:** Implement `content/extractor.js` for Canvas LMS and general web pages.
 3. **Phase 3: Side Panel UI & Design System:** Implement `sidepanel/index.html`, `sidepanel.css` (using `DESIGN.md` tokens), and interactive state rendering.
-4. **Phase 4: LLM Audit Engine & Prompts:** Implement `background/service-worker.js` with structured Gemini 3.7 Flash prompting and JSON schema enforcement.
+4. **Phase 4: LLM Audit Engine & Prompts:** Implement `background/service-worker.js` with structured Gemini 3.7 Flash prompting and JSON schema enforcement. <!-- IDSTACK_CLI_LEAK_ALLOW -->
 5. **Phase 5: Copy Actions & Stakeholder Report Generator:** Wire 1-click clipboard revision copying and standalone HTML report generation.
 6. **Phase 6: Testing & Verification:** Verify on real Canvas pages and web syllabi.
diff --git a/extension/background/canvas-crawler.cjs b/extension/background/canvas-crawler.cjs
deleted file mode 100644
index 23529c1..0000000
--- a/extension/background/canvas-crawler.cjs
+++ /dev/null
@@ -1,59 +0,0 @@
-function stripHtml(html) {
-  if (!html) return '';
-  return html
-    .replace(/<style[^>]*>[\s\S]*?<\/style>/gi, '')
-    .replace(/<script[^>]*>[\s\S]*?<\/script>/gi, '')
-    .replace(/<[^>]+>/g, ' ')
-    .replace(/&nbsp;/gi, ' ')
-    .replace(/&amp;/gi, '&')
-    .replace(/&lt;/gi, '<')
-    .replace(/&gt;/gi, '>')
-    .replace(/&quot;/gi, '"')
-    .replace(/\s+/g, ' ')
-    .trim();
-}
-
-async function crawlCanvasCourse(origin, courseId, fetchImpl = fetch) {
-  if (!origin || !courseId) {
-    throw new Error('Canvas origin and courseId are required for course crawling.');
-  }
-
-  // 1. Fetch Course details & syllabus
-  const courseUrl = `${origin}/api/v1/courses/${courseId}?include[]=syllabus_body`;
-  const courseRes = await fetchImpl(courseUrl, { credentials: 'include' });
-  if (!courseRes.ok) {
-    throw new Error(`Failed to fetch Canvas course info (${courseRes.status})`);
-  }
-  const courseJson = await courseRes.json();
-
-  // 2. Fetch Assignments list (up to 50)
-  const assignmentsUrl = `${origin}/api/v1/courses/${courseId}/assignments?per_page=50`;
-  let assignments = [];
-  try {
-    const assignRes = await fetchImpl(assignmentsUrl, { credentials: 'include' });
-    if (assignRes.ok) {
-      const assignJson = await assignRes.json();
-      if (Array.isArray(assignJson)) {
-        assignments = assignJson.map((a) => ({
-          title: a.name || 'Untitled Assignment',
-          description: stripHtml(a.description || '').slice(0, 1000),
-          points: a.points_possible || 0,
-          dueAt: a.due_at || null
-        }));
-      }
-    }
-  } catch (e) {
-    console.warn('Could not fetch assignments list:', e);
-  }
-
-  return {
-    title: courseJson.name || courseJson.course_code || 'Canvas Course',
-    syllabus: stripHtml(courseJson.syllabus_body || ''),
-    assignments
-  };
-}
-
-module.exports = {
-  stripHtml,
-  crawlCanvasCourse
-};
diff --git a/extension/background/canvas-crawler.js b/extension/background/canvas-crawler.js
index 0c48f1a..6425b78 100644
--- a/extension/background/canvas-crawler.js
+++ b/extension/background/canvas-crawler.js
@@ -13,6 +13,9 @@ export function stripHtml(html) {
     .trim();
 }
 
+// 10 pages of 50 bounds the crawl at 500 assignments.
+const MAX_ASSIGNMENT_PAGES = 10;
+
 export async function crawlCanvasCourse(origin, courseId, fetchImpl = fetch) {
   if (!origin || !courseId) {
     throw new Error('Canvas origin and courseId are required for course crawling.');
@@ -26,24 +29,28 @@ export async function crawlCanvasCourse(origin, courseId, fetchImpl = fetch) {
   }
   const courseJson = await courseRes.json();
 
-  // 2. Fetch Assignments list (up to 50)
-  const assignmentsUrl = `${origin}/api/v1/courses/${courseId}/assignments?per_page=50`;
-  let assignments = [];
-  try {
+  // 2. Fetch every page of the Assignments list (Canvas paginates via the Link
+  // header). Any failure is fatal: an alignment audit run on a partial or empty
+  // list tells the instructor the course has no assessments.
+  const assignments = [];
+  let assignmentsUrl = `${origin}/api/v1/courses/${courseId}/assignments?per_page=50`;
+  for (let page = 1; assignmentsUrl; page++) {
+    if (page > MAX_ASSIGNMENT_PAGES) {
+      throw new Error(`This course has more than ${assignments.length} assignments, more than the full-course audit can read. Try auditing individual assignment pages instead.`);
+    }
     const assignRes = await fetchImpl(assignmentsUrl, { credentials: 'include' });
-    if (assignRes.ok) {
-      const assignJson = await assignRes.json();
-      if (Array.isArray(assignJson)) {
-        assignments = assignJson.map((a) => ({
-          title: a.name || 'Untitled Assignment',
-          description: stripHtml(a.description || '').slice(0, 1000),
-          points: a.points_possible || 0,
-          dueAt: a.due_at || null
-        }));
-      }
+    if (!assignRes.ok) {
+      throw new Error(`Failed to fetch Canvas assignments (${assignRes.status})`);
     }
-  } catch (e) {
-    console.warn('Could not fetch assignments list:', e);
+    const assignJson = await assignRes.json();
+    assignments.push(...assignJson.map((a) => ({
+      title: a.name || 'Untitled Assignment',
+      description: stripHtml(a.description || '').slice(0, 1000),
+      points: a.points_possible || 0,
+      dueAt: a.due_at || null
+    })));
+    const next = /<([^>]+)>;\s*rel="next"/.exec(assignRes.headers.get('Link') || '');
+    assignmentsUrl = next ? next[1] : null;
   }
 
   return {
diff --git a/extension/background/parser-helper.cjs b/extension/background/parser-helper.cjs
deleted file mode 100644
index b98b518..0000000
--- a/extension/background/parser-helper.cjs
+++ /dev/null
@@ -1,128 +0,0 @@
-/**
- * Helper to clean and parse JSON responses from LLM API.
- */
-function cleanJsonResponse(rawText) {
-  let cleaned = rawText.trim();
-  if (cleaned.startsWith('```json')) cleaned = cleaned.replace(/^```json/, '');
-  if (cleaned.startsWith('```')) cleaned = cleaned.replace(/^```/, '');
-  if (cleaned.endsWith('```')) cleaned = cleaned.replace(/```$/, '');
-  return JSON.parse(cleaned.trim());
-}
-
-function parseAuditResponse(rawText) {
-  return cleanJsonResponse(rawText);
-}
-
-function getDemoAuditResult(payload = {}) {
-  const title = (payload && payload.title) ? payload.title : 'Course Material';
-  const pageType = (payload && payload.pageType) ? payload.pageType : 'Assignment / Syllabus';
-
-  return {
-    summary: {
-      bloomsLevel: 'Analyze (Level 4)',
-      alignmentScore: 'Moderate (78%)',
-      keyTakeaway: `[Demo Mode] Sample audit for "${title}" (${pageType}). Entering an API key in Settings unlocks live audits on any page.`
-    },
-    findings: [
-      {
-        severity: 'warning',
-        tier: 'T1',
-        citation: '[Assessment-8] Formative Feedback Matrix',
-        observation: 'Assessment rubric lacks descriptive performance criteria for intermediate mastery levels.',
-        evidence: 'Wisniewski et al. (2020) [T1 meta-analysis, d=0.48]: Elaborated feedback and rubric transparency significantly boost student self-regulation and achievement.',
-        recommendation: 'Add concrete performance descriptors and milestone criteria for each grade band instead of generic labels.'
-      },
-      {
-        severity: 'critical',
-        tier: 'T2',
-        citation: '[Alignment-3] Direct Constructive Alignment',
-        observation: 'Learning objectives target analytical synthesis, but evaluation instruments only test lower-order recall.',
-        evidence: 'Biggs (1996) & Liou et al. (2023) [T2 controlled trial]: Constructive misalignment between stated objectives and assessment formats leads to superficial learning strategies.',
-        recommendation: 'Incorporate authentic problem-solving prompts and case analysis rather than purely multiple-choice recall questions.'
-      },
-      {
-        severity: 'info',
-        tier: 'T1',
-        citation: '[Cognitive-2] Cognitive Load & Chunking',
-        observation: 'Task instructions present multiple complex requirements in a single unsegmented block.',
-        evidence: 'Costley et al. (2023) [T1 meta-analysis]: Segmenting complex instructional tasks into structured sequential phases reduces extraneous cognitive load.',
-        recommendation: 'Format multi-step assignment guidelines into sequenced checklists or distinct milestone stages.'
-      }
-    ],
-    improvedDraft: {
-      title: 'Improved Rubric Draft (idstack Recommended)',
-      content: `### Revised Task & Assessment Rubric for: ${title}
-
-> **Note:** This is a sample evidence-based draft generated in demo mode. Entering an API key in Settings unlocks live audits on any page.
-
-#### Analytical Task Prompt
-Analyze the core case scenario and formulate a structured recommendation addressing:
-1. Primary contributing factors identified in the evidence base.
-2. Direct trade-offs between proposed intervention strategies.
-3. A measurable evaluation plan for validating outcomes.
-
-#### Evaluation Rubric (Elaborated Criteria)
-| Criterion | Exemplary (Proficient) | Developing | Novice |
-| :--- | :--- | :--- | :--- |
-| **Evidence Application [T1]** | Synthesizes 3+ relevant empirical sources with clear justification. | References 1-2 sources with partial justification. | Makes claims without empirical citations. |
-| **Analytical Rigor [T2]** | Rigorously evaluates trade-offs and alternative hypotheses. | Identifies trade-offs but lacks systematic comparison. | Describes facts without evaluating trade-offs. |
-| **Actionable Strategy [T1]** | Proposes concrete, measurable milestones with validation metrics. | Proposes general steps without measurable metrics. | Vague or non-actionable suggestions. |`
-    }
-  };
-}
-
-function getDemoCourseAuditResult(payload = {}) {
-  const title = payload.title || 'Sample Canvas Course';
-  return {
-    summary: {
-      bloomsLevel: 'Analyze & Evaluate (Levels 4-5)',
-      alignmentScore: 'Moderate (74%)',
-      keyTakeaway: `[Course-Level Demo] Full-course audit for "${title}". Alignment gaps identified between week 1-4 recall quizzes and week 12 analytical capstone.`
-    },
-    findings: [
-      {
-        severity: 'critical',
-        tier: 'T2',
-        citation: '[Alignment-3] Direct Constructive Alignment',
-        observation: 'Modules 1-6 assess solely lower-order factual recall, while the final course project demands high-order synthesis without intermediate scaffolding.',
-        evidence: 'Biggs (1996) & Liou et al. (2023) [T2 controlled trial]: Abrupt jumps in cognitive demand without progressive assessment scaffolding increase failure rates.',
-        recommendation: 'Introduce mid-semester milestone case studies in Module 4 to bridge the gap between quizzes and the final capstone.'
-      },
-      {
-        severity: 'warning',
-        tier: 'T1',
-        citation: '[Cognitive-2] Cognitive Load & Spaced Practice',
-        observation: 'Major assignment deadlines are clustered in Week 14-15 with no spaced formative checkpoints.',
-        evidence: 'Carpenter et al. (2022) [T1 meta-analysis, d=0.61]: Distributing assessments across spaced intervals produces significantly higher long-term retention.',
-        recommendation: 'Redistribute submission checkpoints into 3 progressive deliverables across weeks 6, 10, and 14.'
-      },
-      {
-        severity: 'info',
-        tier: 'T1',
-        citation: '[Assessment-8] Formative Rubric Transparency',
-        observation: 'Syllabus grading policy lacks explicit performance criteria for collaborative group deliverables.',
-        evidence: 'Wisniewski et al. (2020) [T1 meta-analysis]: Pre-distribution of analytic rubrics with milestone criteria boosts student self-regulation and achievement.',
-        recommendation: 'Publish the multi-tier analytic grading rubric during the initial module launch.'
-      }
-    ],
-    improvedDraft: {
-      title: 'Course-Wide Constructive Alignment Matrix',
-      content: `### Course Alignment Matrix: ${title}
-
-| Week / Module | Intended Learning Outcome | Formative Checkpoint [T1] | Summative Assessment [T2] |
-| :--- | :--- | :--- | :--- |
-| **Weeks 1-3** | Foundational Cell Structure | Spaced Knowledge Check (10 pts) | Module 1 Synthesis Quiz |
-| **Weeks 4-7** | Enzyme Kinetics & Modeling | Case Problem Milestone 1 [T2] | Lab Protocol Analysis |
-| **Weeks 8-11** | Experimental Troubleshooting | Peer Review Protocol [T1] | Milestone 2 Experimental Draft |
-| **Weeks 12-15**| Autonomous Investigation | Scaffolded Capstone Consult | Final Research Capstone |`
-    }
-  };
-}
-
-module.exports = {
-  cleanJsonResponse,
-  parseAuditResponse,
-  getDemoAuditResult,
-  getDemoCourseAuditResult
-};
-
diff --git a/extension/background/parser-helper.js b/extension/background/parser-helper.js
index 9890569..dbf8c91 100644
--- a/extension/background/parser-helper.js
+++ b/extension/background/parser-helper.js
@@ -34,18 +34,18 @@ export function getDemoAuditResult(payload = {}) {
       },
       {
         severity: 'critical',
-        tier: 'T2',
-        citation: '[Alignment-3] Direct Constructive Alignment',
+        tier: 'T5',
+        citation: '[Alignment-1] Direct Constructive Alignment',
         observation: 'Learning objectives target analytical synthesis, but evaluation instruments only test lower-order recall.',
-        evidence: 'Biggs (1996) & Liou et al. (2023) [T2 controlled trial]: Constructive misalignment between stated objectives and assessment formats leads to superficial learning strategies.',
+        evidence: 'Biggs (1996) [T5]: Constructive misalignment between stated objectives and assessment formats leads to superficial learning strategies.',
         recommendation: 'Incorporate authentic problem-solving prompts and case analysis rather than purely multiple-choice recall questions.'
       },
       {
         severity: 'info',
         tier: 'T1',
-        citation: '[Cognitive-2] Cognitive Load & Chunking',
+        citation: '[CogLoad-1] Cognitive Load & Chunking',
         observation: 'Task instructions present multiple complex requirements in a single unsegmented block.',
-        evidence: 'Costley et al. (2023) [T1 meta-analysis]: Segmenting complex instructional tasks into structured sequential phases reduces extraneous cognitive load.',
+        evidence: 'Costley et al. (2023) [T1]: Segmenting complex instructional tasks into structured sequential phases reduces extraneous cognitive load.',
         recommendation: 'Format multi-step assignment guidelines into sequenced checklists or distinct milestone stages.'
       }
     ],
@@ -82,18 +82,18 @@ export function getDemoCourseAuditResult(payload = {}) {
     findings: [
       {
         severity: 'critical',
-        tier: 'T2',
-        citation: '[Alignment-3] Direct Constructive Alignment',
+        tier: 'T5',
+        citation: '[Alignment-1] Direct Constructive Alignment',
         observation: 'Modules 1-6 assess solely lower-order factual recall, while the final course project demands high-order synthesis without intermediate scaffolding.',
-        evidence: 'Biggs (1996) & Liou et al. (2023) [T2 controlled trial]: Abrupt jumps in cognitive demand without progressive assessment scaffolding increase failure rates.',
+        evidence: 'Biggs (1996) [T5]: Constructive alignment requires every assessment to engage the cognitive level its learning outcome names.',
         recommendation: 'Introduce mid-semester milestone case studies in Module 4 to bridge the gap between quizzes and the final capstone.'
       },
       {
         severity: 'warning',
         tier: 'T1',
-        citation: '[Cognitive-2] Cognitive Load & Spaced Practice',
+        citation: '[CogLoad-6] Cognitive Load & Spaced Practice',
         observation: 'Major assignment deadlines are clustered in Week 14-15 with no spaced formative checkpoints.',
-        evidence: 'Carpenter et al. (2022) [T1 meta-analysis, d=0.61]: Distributing assessments across spaced intervals produces significantly higher long-term retention.',
+        evidence: 'Chen et al. (2018) [T1]: Distributing practice across time produces stronger long-term retention than massing the same content into a single session.',
         recommendation: 'Redistribute submission checkpoints into 3 progressive deliverables across weeks 6, 10, and 14.'
       },
       {
diff --git a/extension/background/service-worker.js b/extension/background/service-worker.js
index 16093b6..58050d5 100644
--- a/extension/background/service-worker.js
+++ b/extension/background/service-worker.js
@@ -4,26 +4,46 @@ import { cleanJsonResponse, getDemoAuditResult, getDemoCourseAuditResult } from
 import { crawlCanvasCourse } from './canvas-crawler.js';
 import { verifyFindingsWithConsensus } from '../shared/consensus-client.js';
 
-// Setup side panel behavior on action click
+const LLM_TIMEOUT_MS = 25000;   // Chrome kills an extension SW whose fetch response takes >30 s
+// Fewer words than this means an unreadable page (browser page, PDF viewer,
+// a single Modules item title); the model would return an audit of nothing.
+const MIN_AUDIT_WORDS = 10;
+
+// Open the panel from action.onClicked, not openPanelOnActionClick: Chrome routes a
+// side-panel-toggle click past its activeTab grant (crbug.com/40904917), and activeTab
+// is how the panel reads pages now that no content script is declared. v1.0.0 stored
+// `true` in the extension's prefs, so set it back to false explicitly.
 if (typeof chrome !== 'undefined' && chrome.sidePanel && chrome.sidePanel.setPanelBehavior) {
-  chrome.sidePanel.setPanelBehavior({ openPanelOnActionClick: true }).catch((err) => console.error(err));
+  chrome.sidePanel.setPanelBehavior({ openPanelOnActionClick: false }).catch((err) => console.error(err));
+}
+if (typeof chrome !== 'undefined' && chrome.action && chrome.action.onClicked) {
+  chrome.action.onClicked.addListener((tab) => {
+    // No await before open(): it needs the click's user gesture.
+    chrome.sidePanel.open({ windowId: tab.windowId });
+    // An open panel re-reads the tab it can now access; if none is open yet, nothing listens.
+    chrome.runtime.sendMessage({ action: 'TAB_ACCESS_GRANTED' }).catch(() => {});
+  });
 }
 
 async function callLlmApi(apiKey, prompt) {
-  const url = `https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent`; // IDSTACK_CLI_LEAK_ALLOW
+  const url = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent'; // IDSTACK_CLI_LEAK_ALLOW
+  const signal = AbortSignal.timeout(LLM_TIMEOUT_MS);
   const response = await fetch(url, {
     method: 'POST',
-    headers: {
-      'Content-Type': 'application/json',
-      'x-goog-api-key': apiKey
-    },
+    headers: { 'Content-Type': 'application/json', 'x-goog-api-key': apiKey },
     body: JSON.stringify({
       contents: [{ parts: [{ text: prompt }] }],
       generationConfig: {
         responseMimeType: 'application/json',
         temperature: 0.2
       }
-    })
+    }),
+    signal
+  }).catch((err) => {
+    // Chrome < 124 rejects a timed-out fetch with AbortError, so check the signal, not err.name.
+    throw signal.aborted
+      ? new Error(`The AI service did not respond within ${LLM_TIMEOUT_MS / 1000} seconds. Please retry the audit.`)
+      : err;
   });
 
   if (!response.ok) {
@@ -35,14 +55,29 @@ async function callLlmApi(apiKey, prompt) {
   const rawText = data.candidates?.[0]?.content?.parts?.[0]?.text;
   if (!rawText) throw new Error('Empty response from AI model.');
 
-  return cleanJsonResponse(rawText);
+  const result = cleanJsonResponse(rawText);
+  // The renderer and dossier compiler read every finding and call string methods
+  // on tier and severity; refuse a shape they cannot handle before it is saved.
+  const optionalString = (v) => v == null || typeof v === 'string';
+  if (!Array.isArray(result?.findings) || !result.findings.every((f) => f && typeof f === 'object' && optionalString(f.tier) && optionalString(f.severity))) {
+    throw new Error('The AI response did not match the audit format. Please retry the audit.');
+  }
+  return result;
 }
 
 if (typeof chrome !== 'undefined' && chrome.runtime && chrome.runtime.onMessage) {
   chrome.runtime.onMessage.addListener((request, sender, sendResponse) => {
+    // Only the extension's own pages (the side panel) may start audits. A content-script
+    // context shares sender.id, so compare the sender's URL with our origin instead.
+    if (!String(sender.url || '').startsWith(chrome.runtime.getURL(''))) return;
     if (request.action === 'RUN_AUDIT') {
       (async () => {
         try {
+          const wordCount = String(request.payload?.content || '').split(/\s+/).filter(Boolean).length;
+          if (wordCount < MIN_AUDIT_WORDS) {
+            // The extractor and side panel explain pages they could not or would not read.
+            throw new Error(request.payload?.emptyReason || `This page has too little readable text to audit (${wordCount} words; at least ${MIN_AUDIT_WORDS} needed). Browser pages, PDF viewers, and image-only pages expose no text to idstack.`);
+          }
           const settings = await getSettings();
           
           let auditResult;
@@ -77,7 +112,17 @@ if (typeof chrome !== 'undefined' && chrome.runtime && chrome.runtime.onMessage)
       (async () => {
         try {
           const { origin, courseId } = request.payload || {};
+          // origin is spliced into a credentialed fetch URL: accept only a bare https
+          // origin and a numeric course id, so no path or query can ride along.
+          let target = null;
+          try { target = new URL(origin); } catch (e) { /* not a URL */ }
+          if (!target || target.protocol !== 'https:' || target.origin !== origin || !/^\d+$/.test(courseId)) {
+            throw new Error('Invalid Canvas course target.');
+          }
           const courseData = await crawlCanvasCourse(origin, courseId);
+          if (!courseData.syllabus.trim() && courseData.assignments.length === 0) {
+            throw new Error('This course has no syllabus text and no assignments for idstack to audit.');
+          }
           const settings = await getSettings();
 
           let auditResult;
diff --git a/extension/content/extractor-core.cjs b/extension/content/extractor-core.cjs
deleted file mode 100644
index 0256253..0000000
--- a/extension/content/extractor-core.cjs
+++ /dev/null
@@ -1,8 +0,0 @@
-const { detectPageType, extractContentFromDOM, extractPageContent, detectCourseContext } = require('./extractor.js');
-
-module.exports = {
-  detectPageType,
-  extractContentFromDOM,
-  extractPageContent,
-  detectCourseContext
-};
diff --git a/extension/content/extractor-core.js b/extension/content/extractor-core.js
index 35af7e2..799f60d 100644
--- a/extension/content/extractor-core.js
+++ b/extension/content/extractor-core.js
@@ -1,49 +1,4 @@
-export function detectPageType(url, document) {
-  if (url.includes('instructure.com') || url.includes('/courses/')) {
-    if (document.querySelector('#assignment_show')) return 'Canvas Assignment';
-    if (document.querySelector('#syllabusContainer') || url.includes('/assignments/syllabus')) return 'Canvas Syllabus';
-    if (document.querySelector('#modules') || url.includes('/modules')) return 'Canvas Modules';
-    if (document.querySelector('#rubrics')) return 'Canvas Rubric';
-    return 'Canvas LMS Page';
-  }
-  if (url.includes('docs.google.com/document')) return 'Google Doc Syllabus';
-  return 'Web Syllabus / Course Page';
-}
-
-export function extractContentFromDOM(document, url = (typeof window !== 'undefined' && window.location ? window.location.href : '')) {
-  const pageType = detectPageType(url, document);
-  let title = (document && document.title) || 'Course Document';
-  let content = '';
-
-  if (pageType.startsWith('Canvas')) {
-    const heading = document.querySelector('#assignment_show .title, .page-title, h1');
-    if (heading) title = (heading.innerText || heading.textContent || '').trim();
-
-    const mainBody = document.querySelector('#assignment_show .description.user_content, .show-content.user_content, #syllabusContainer, .module-item-title');
-    content = mainBody ? (mainBody.innerText || mainBody.textContent || '').trim() : (document.body ? (document.body.innerText || document.body.textContent || '').trim() : '');
-  } else if (pageType === 'Google Doc Syllabus') {
-    const kixApp = document.querySelector('.kix-appview-editor');
-    content = kixApp ? (kixApp.innerText || kixApp.textContent || '').trim() : (document.body ? (document.body.innerText || document.body.textContent || '').trim() : '');
-  } else {
-    // General web page extraction
-    const article = document.querySelector('main, article, [role="main"]');
-    content = article ? (article.innerText || article.textContent || '').trim() : (document.body ? (document.body.innerText || document.body.textContent || '').trim() : '');
-  }
-
-  return {
-    url,
-    title,
-    pageType,
-    content: content.slice(0, 15000),
-    wordCount: (content.match(/\S+/g) || []).length
-  };
-}
-
-export function extractPageContent() {
-  const url = typeof window !== 'undefined' && window.location ? window.location.href : '';
-  return extractContentFromDOM(document, url);
-}
-
+// Canvas course-root detection for the side panel. The content script does not use it.
 export function detectCourseContext(url, docTitle) {
   if (!url) return { isCourseRoot: false, courseId: null, origin: null };
   try {
diff --git a/extension/content/extractor.js b/extension/content/extractor.js
index d59b9f5..ca50fff 100644
--- a/extension/content/extractor.js
+++ b/extension/content/extractor.js
@@ -2,10 +2,10 @@
  * Injected content script to extract course content from Canvas, Google Docs, and web pages.
  */
 function detectPageType(url, document) {
-  if (url.includes('instructure.com') || url.includes('/courses/')) {
+  if (url.includes('instructure.com') || /\/courses\/\d+(?:[/?#]|$)/.test(url)) {
     if (document.querySelector('#assignment_show')) return 'Canvas Assignment';
     if (document.querySelector('#syllabusContainer') || url.includes('/assignments/syllabus')) return 'Canvas Syllabus';
-    if (document.querySelector('#modules') || url.includes('/modules')) return 'Canvas Modules';
+    if (document.querySelector('#context_modules, #modules') || url.includes('/modules')) return 'Canvas Modules';
     if (document.querySelector('#rubrics')) return 'Canvas Rubric';
     return 'Canvas LMS Page';
   }
@@ -14,6 +14,23 @@ function detectPageType(url, document) {
 }
 
 function extractContentFromDOM(document, url = (typeof window !== 'undefined' && window.location ? window.location.href : '')) {
+  // Canvas views that show student records: gradebook and SpeedGrader, grades, People
+  // (course and account level), groups, discussion replies, submissions, Inbox. Checked on
+  // every URL, because Canvas on a custom domain is only detected as Canvas under /courses/N.
+  // Matched on the URL string: url can be '', and new URL('') throws. Function-scoped on
+  // purpose: a top-level const would throw when the file is re-injected into the same world.
+  const studentRecordUrl = /^https?:\/\/[^/?#]+\/(?:courses\/\d+\/(?:gradebook|grades|users|groups|discussion_topics|assignments\/\d+\/submissions)|conversations|groups\/\d+|users\/\d+|accounts\/\d+\/users)(?:[/?#]|$)/;
+  if (studentRecordUrl.test(url)) {
+    return {
+      url,
+      title: 'Student records page',
+      pageType: 'Canvas Student Records',
+      content: '',
+      wordCount: 0,
+      emptyReason: 'idstack does not read Canvas pages that show student records (grades, People, Inbox, discussions, groups, submissions). Consider auditing an assignment, page, syllabus, rubric, or Modules view instead.'
+    };
+  }
+
   const pageType = detectPageType(url, document);
   let title = (document && document.title) || 'Course Document';
   let content = '';
@@ -22,8 +39,17 @@ function extractContentFromDOM(document, url = (typeof window !== 'undefined' &&
     const heading = document.querySelector('#assignment_show .title, .page-title, h1');
     if (heading) title = (heading.innerText || heading.textContent || '').trim();
 
-    const mainBody = document.querySelector('#assignment_show .description.user_content, .show-content.user_content, #syllabusContainer, .module-item-title');
-    content = mainBody ? (mainBody.innerText || mainBody.textContent || '').trim() : (document.body ? (document.body.innerText || document.body.textContent || '').trim() : '');
+    if (pageType === 'Canvas Modules') {
+      // Every item title; querySelector alone returns only the first one.
+      content = Array.from(document.querySelectorAll('.module-item-title'))
+        .map((el) => (el.innerText || el.textContent || '').trim())
+        .filter(Boolean)
+        .join('\n');
+    } else {
+      const mainBody = document.querySelector('#assignment_show .description.user_content, .show-content.user_content, #syllabusContainer, #rubrics');
+      // No document.body fallback on Canvas: an unmatched view can hold names, grades, or posts.
+      content = mainBody ? (mainBody.innerText || mainBody.textContent || '').trim() : '';
+    }
   } else if (pageType === 'Google Doc Syllabus') {
     const kixApp = document.querySelector('.kix-appview-editor');
     content = kixApp ? (kixApp.innerText || kixApp.textContent || '').trim() : (document.body ? (document.body.innerText || document.body.textContent || '').trim() : '');
@@ -42,43 +68,40 @@ function extractContentFromDOM(document, url = (typeof window !== 'undefined' &&
   };
 }
 
-function extractPageContent() {
+// Async, so the injection's completion value is a Promise; Chrome awaits it and the
+// side panel still receives the payload.
+async function extractPageContent() {
   const url = typeof window !== 'undefined' && window.location ? window.location.href : '';
-  return extractContentFromDOM(document, url);
-}
-
-function detectCourseContext(url, docTitle) {
-  if (!url) return { isCourseRoot: false, courseId: null, origin: null };
+  const data = extractContentFromDOM(document, url);
+  // Google Docs draws its text on <canvas>, so the editor DOM holds none of it. Read the
+  // plain-text export with the user's own session instead (same origin, no new permission).
+  // Keep fetch's default credentials mode: the export redirects to a host that answers
+  // ACAO: *, which a credentialed request would reject. Published /d/e/ docs are skipped:
+  // their text is in the DOM.
+  const doc = url.match(/^https:\/\/docs\.google\.com\/document\/(?:u\/\d+\/)?d\/(?!e\/)[\w-]+/);
+  if (!doc) return data;
   try {
-    const parsedUrl = new URL(url);
-    const match = parsedUrl.pathname.match(/\/courses\/(\d+)(?:\/(?:modules)?)?\/?$/);
-    const anyCourseMatch = parsedUrl.pathname.match(/\/courses\/(\d+)/);
-    return {
-      isCourseRoot: !!match,
-      courseId: anyCourseMatch ? anyCourseMatch[1] : null,
-      origin: parsedUrl.origin
-    };
+    const res = await fetch(`${doc[0]}/export?format=txt`, { signal: AbortSignal.timeout(10000) });
+    // Signed out, the export answers 200 with an HTML sign-in page.
+    if (!res.ok || !(res.headers.get('content-type') || '').startsWith('text/plain')) throw new Error(String(res.status));
+    const text = (await res.text()).trim();
+    return { ...data, content: text.slice(0, 15000), wordCount: text.split(/\s+/).filter(Boolean).length };
   } catch (e) {
-    return { isCourseRoot: false, courseId: null, origin: null };
+    // Offline, timed out, refused, or not plain text.
+    return { ...data, content: '', wordCount: 0, emptyReason: 'idstack could not read the text of this Google Doc. If downloading is turned off for it, consider asking the owner to allow downloads.' };
   }
 }
 
-// Listen for messages from Side Panel
-if (typeof chrome !== 'undefined' && chrome.runtime && chrome.runtime.onMessage) {
-  chrome.runtime.onMessage.addListener((request, sender, sendResponse) => {
-    if (request.action === 'EXTRACT_CONTENT') {
-      const data = extractPageContent();
-      sendResponse(data);
-    }
-    return true;
-  });
-}
-
+// Injected with chrome.scripting.executeScript({ files }): the value of the final
+// statement is the InjectionResult the side panel reads, so extractPageContent()
+// must stay last. Re-injected on every refresh into the same isolated world, so no
+// top-level let/const/class (redeclaring them throws).
 if (typeof module !== 'undefined' && module.exports) {
   module.exports = {
     detectPageType,
     extractContentFromDOM,
-    extractPageContent,
-    detectCourseContext
+    extractPageContent
   };
+} else {
+  extractPageContent();
 }
diff --git a/extension/manifest.json b/extension/manifest.json
index bea6910..7949678 100644
--- a/extension/manifest.json
+++ b/extension/manifest.json
@@ -1,7 +1,8 @@
 {
   "manifest_version": 3,
   "name": "idstack — Evidence-Based Course Design",
-  "version": "1.0.0",
+  "version": "1.1.0",
+  "minimum_chrome_version": "116",
   "description": "Evidence-based instructional design co-pilot for Canvas LMS, Google Docs, and course web pages.",
   "homepage_url": "https://idstack.org",
   "icons": {
@@ -19,17 +20,6 @@
     "service_worker": "background/service-worker.js",
     "type": "module"
   },
-  "content_scripts": [
-    {
-      "matches": [
-        "*://*.instructure.com/*",
-        "*://docs.google.com/document/*",
-        "<all_urls>"
-      ],
-      "js": ["content/extractor.js"],
-      "run_at": "document_idle"
-    }
-  ],
   "permissions": [
     "sidePanel",
     "storage",
@@ -40,5 +30,8 @@
     "https://generativelanguage.googleapis.com/*",
     "*://*.instructure.com/*",
     "https://api.consensus.app/*"
+  ],
+  "optional_host_permissions": [
+    "https://*/*"
   ]
 }
diff --git a/extension/shared/consensus-client.cjs b/extension/shared/consensus-client.cjs
deleted file mode 100644
index 7b180ab..0000000
--- a/extension/shared/consensus-client.cjs
+++ /dev/null
@@ -1,352 +0,0 @@
-/**
- * Consensus evidence client and background QA pass for Chrome extension (CommonJS).
- * Provides neuromyth auto-correction, cache checking, and Consensus API enrichment.
- */
-
-const NEUROMYTH_PATTERNS = [
-  {
-    type: 'learning_styles',
-    pattern: /\b(learning\s+styles?|visual\s+(?:vs\.?|and|or)\s+auditory\s+learn|visual\s+learners?|auditory\s+learners?|kinesthetic\s+learners?|meshing\s+hypothesis|cater(?:ing)?\s+to\s+(?:individual\s+)?learning\s+styles?)\b/i
-  },
-  {
-    type: 'hemisphere_learning',
-    pattern: /\b(left[-\s]brain|right[-\s]brain|hemisphere\s+learn)\b/i
-  },
-  {
-    type: 'ten_percent_brain',
-    pattern: /\b(10%\s+of\s+(?:the|their)?\s*brain)\b/i
-  },
-  {
-    type: 'dales_cone_percentages',
-    pattern: /\b(?:remember|retain)\s+\d+%\s+of\s+what\s+(?:they|we)\s+(?:read|hear|see)\b/i
-  }
-];
-
-async function hashClaim(text) {
-  const norm = (text || '').toLowerCase().trim().replace(/\s+/g, ' ');
-  if (typeof globalThis !== 'undefined' && globalThis.crypto && globalThis.crypto.subtle) {
-    const encoder = new TextEncoder();
-    const data = encoder.encode(norm);
-    const hashBuffer = await globalThis.crypto.subtle.digest('SHA-256', data);
-    return Array.from(new Uint8Array(hashBuffer))
-      .map((b) => b.toString(16).padStart(2, '0'))
-      .join('');
-  }
-  if (typeof require === 'function') {
-    try {
-      const nodeCrypto = require('crypto');
-      return nodeCrypto.createHash('sha256').update(norm).digest('hex');
-    } catch {
-      // ignore
-    }
-  }
-  let hash = 0;
-  for (let i = 0; i < norm.length; i++) {
-    hash = ((hash << 5) - hash) + norm.charCodeAt(i);
-    hash |= 0;
-  }
-  return Math.abs(hash).toString(16);
-}
-
-function detectNeuromyth(finding) {
-  if (!finding) return null;
-  const combined = [
-    finding.recommendation || '',
-    finding.evidence || '',
-    finding.observation || '',
-    finding.citation || '',
-    finding.claim || ''
-  ].join(' ');
-  for (const { type, pattern } of NEUROMYTH_PATTERNS) {
-    if (pattern.test(combined)) {
-      return type;
-    }
-  }
-  return null;
-}
-
-function autoCorrectNeuromyth(finding, mythType) {
-  if (!finding) return finding;
-  finding.auto_corrected = true;
-  finding.qa_verified = true;
-
-  if (mythType === 'learning_styles') {
-    finding.recommendation =
-      'Provide multimodal presentation and dual coding (combining complementary visual representations ' +
-      'and verbal/textual explanations for all learners) rather than attempting to segregate ' +
-      'or adapt materials to putative learning styles.';
-    const corrEv =
-      "[Auto-Corrected Neuromyth] The 'learning styles' (meshing hypothesis) lacks empirical support " +
-      '(Pashler et al., 2008; Rogowsky et al., 2015). Empirical evidence supports multimodal ' +
-      'presentation and dual coding (Mayer, 2021; Paivio, 1986) for all learners.';
-    finding.evidence = finding.evidence ? `${finding.evidence} ${corrEv}` : corrEv;
-    const obsNote = ' [Note: Differentiating by sensory learning styles is a recognized neuromyth; multimodal presentation should be applied instead.]';
-    finding.observation = `${finding.observation || ''}${obsNote}`.trim();
-  } else if (mythType === 'hemisphere_learning') {
-    finding.recommendation =
-      'Provide multimodal presentation and integrated cognitive activities rather than ' +
-      'attempting to cater to left-brain or right-brain learners.';
-    const corrEv =
-      '[Auto-Corrected Neuromyth] Hemispheric specialization for learning is an empirically ' +
-      'unsupported neuromyth. Cognitive architecture operates through integrated networks; ' +
-      'multimodal dual coding benefits all learners.';
-    finding.evidence = finding.evidence ? `${finding.evidence} ${corrEv}` : corrEv;
-    const obsNote = ' [Note: Left/right brain learning categorization is a neuromyth; integrated multimodal tasks should be used.]';
-    finding.observation = `${finding.observation || ''}${obsNote}`.trim();
-  } else if (mythType === 'ten_percent_brain') {
-    finding.recommendation =
-      'Design instructional activities that optimize working memory capacity and cognitive load ' +
-      'rather than relying on 10% brain capacity claims.';
-    const corrEv =
-      '[Auto-Corrected Neuromyth] The 10% brain usage claim is a debunked neuromyth. Instructional ' +
-      'design should be guided by Cognitive Load Theory (Sweller, 1994) and working memory limits.';
-    finding.evidence = finding.evidence ? `${finding.evidence} ${corrEv}` : corrEv;
-    const obsNote = ' [Note: The 10% brain claim is a recognized neuromyth; Cognitive Load Theory should guide design.]';
-    finding.observation = `${finding.observation || ''}${obsNote}`.trim();
-  } else if (mythType === 'dales_cone_percentages') {
-    finding.recommendation =
-      'Structure learning through active retrieval and generative processing rather than ' +
-      'bogus retention percentage hierarchies.';
-    const corrEv =
-      "[Auto-Corrected Neuromyth] Numerical retention claims ('we remember 10% of what we read') " +
-      "are fabricated percentages falsely attributed to Dale's Cone of Experience. Retention " +
-      'depends on retrieval practice and cognitive processing depth (Agarwal, 2019).';
-    finding.evidence = finding.evidence ? `${finding.evidence} ${corrEv}` : corrEv;
-    const obsNote = ' [Note: Retention percentage pyramids are fabricated; active retrieval principles should be used.]';
-    finding.observation = `${finding.observation || ''}${obsNote}`.trim();
-  }
-  return finding;
-}
-
-function formatConsensus(record) {
-  if (!record) return null;
-  if (record.consensus) {
-    return {
-      meter: record.consensus.meter ?? record.consensus.yes_pct ?? 88,
-      totalStudies: record.consensus.totalStudies ?? record.consensus.total_papers ?? 1,
-      verified: true,
-      paperUrl: record.consensus.paperUrl || record.consensus.doi_url || ''
-    };
-  }
-  if (record.consensus_meter) {
-    const topPaper = record.top_papers?.[0];
-    return {
-      meter: record.consensus_meter.yes_pct ?? 88,
-      totalStudies: record.consensus_meter.total_papers ?? (record.top_papers?.length || 1),
-      verified: true,
-      paperUrl: topPaper?.doi_url || topPaper?.url || ''
-    };
-  }
-  return {
-    meter: typeof record.meter === 'number' ? record.meter : 88,
-    totalStudies: typeof record.totalStudies === 'number' ? record.totalStudies : 1,
-    verified: true,
-    paperUrl: record.paperUrl || record.doi_url || ''
-  };
-}
-
-function resolveStorage(storage) {
-  if (storage) return storage;
-  if (typeof chrome !== 'undefined' && chrome.storage && chrome.storage.local) {
-    return chrome.storage.local;
-  }
-  return null;
-}
-
-async function getCachedRecord(storage, cacheKey) {
-  const st = resolveStorage(storage);
-  if (!st) return null;
-  if (typeof st.get === 'function') {
-    return new Promise((resolve) => {
-      let resolved = false;
-      try {
-        const res = st.get([cacheKey], (data) => {
-          if (!resolved) {
-            resolved = true;
-            resolve(data ? data[cacheKey] : null);
-          }
-        });
-        if (res && typeof res.then === 'function') {
-          res
-            .then((data) => {
-              if (!resolved) {
-                resolved = true;
-                resolve(data ? (data[cacheKey] !== undefined ? data[cacheKey] : data) : null);
-              }
-            })
-            .catch(() => {
-              if (!resolved) {
-                resolved = true;
-                resolve(null);
-              }
-            });
-        }
-      } catch {
-        try {
-          const syncRes = st.get(cacheKey);
-          resolve(syncRes ? (syncRes[cacheKey] !== undefined ? syncRes[cacheKey] : syncRes) : null);
-        } catch {
-          resolve(null);
-        }
-      }
-    });
-  }
-  if (typeof st === 'object' && st[cacheKey] !== undefined) {
-    return st[cacheKey];
-  }
-  return null;
-}
-
-async function setCachedRecord(storage, cacheKey, value) {
-  const st = resolveStorage(storage);
-  if (!st) return;
-  if (typeof st.set === 'function') {
-    return new Promise((resolve) => {
-      let resolved = false;
-      try {
-        const res = st.set({ [cacheKey]: value }, () => {
-          if (!resolved) {
-            resolved = true;
-            resolve(true);
-          }
-        });
-        if (res && typeof res.then === 'function') {
-          res
-            .then(() => {
-              if (!resolved) {
-                resolved = true;
-                resolve(true);
-              }
-            })
-            .catch(() => {
-              if (!resolved) {
-                resolved = true;
-                resolve(false);
-              }
-            });
-        }
-      } catch {
-        resolve(false);
-      }
-    });
-  }
-  if (typeof st === 'object') {
-    st[cacheKey] = value;
-  }
-}
-
-function extractClaimText(finding) {
-  return (
-    finding.claim ||
-    finding.evidence ||
-    finding.observation ||
-    finding.recommendation ||
-    finding.citation ||
-    ''
-  ).trim();
-}
-
-/**
- * Verifies findings with Consensus: checks cache, auto-corrects neuromyths,
- * and fetches from Consensus API if API key is provided.
- *
- * @param {Array} findings Array of finding objects
- * @param {string} [apiKey] Consensus API key
- * @param {Object} [storage] Storage instance (defaults to chrome.storage.local)
- * @returns {Promise<Array>} Verified findings array
- */
-async function verifyFindingsWithConsensus(findings, apiKey, storage = null) {
-  if (!Array.isArray(findings)) {
-    return findings;
-  }
-
-  for (const finding of findings) {
-    if (!finding || typeof finding !== 'object') continue;
-
-    // 1. Detect and auto-correct known neuromyths
-    const mythType = detectNeuromyth(finding);
-    if (mythType) {
-      autoCorrectNeuromyth(finding, mythType);
-    }
-
-    // Ensure default tier metadata exists
-    if (!finding.tier) {
-      finding.tier = 'T5';
-    }
-
-    const claimText = extractClaimText(finding);
-    if (!claimText) continue;
-
-    const hash = await hashClaim(claimText);
-    const cacheKey = `consensus_cache_${hash}`;
-
-    // 2. Check storage cache
-    const cached = await getCachedRecord(storage, cacheKey);
-    if (cached) {
-      const consensusData = formatConsensus(cached);
-      if (consensusData) {
-        finding.consensus = consensusData;
-        finding.qa_verified = true;
-        if (consensusData.paperUrl) {
-          finding.doi_url = consensusData.paperUrl;
-        }
-      }
-      continue;
-    }
-
-    // 3. If API key is provided and item is not in cache, query Consensus API
-    if (apiKey && typeof apiKey === 'string' && apiKey.trim()) {
-      try {
-        const cleanQuery = claimText.replace(/[\.\,\;\:\!\?]+$/, '').trim();
-        const res = await fetch('https://api.consensus.app/v1/search', {
-          method: 'POST',
-          headers: {
-            'Content-Type': 'application/json',
-            'Authorization': `Bearer ${apiKey.trim()}`
-          },
-          body: JSON.stringify({ query: cleanQuery || claimText, limit: 5 })
-        });
-
-        if (res.ok) {
-          const data = await res.json();
-          const papers = data.papers || data.results || [];
-          const topPaper = papers[0];
-          const paperUrl = topPaper?.doi_url || topPaper?.url || (topPaper?.doi ? `https://doi.org/${topPaper.doi}` : '');
-          const meter = data.consensus_meter?.yes_pct ?? (papers.length > 0 ? 85 : 50);
-          const totalStudies = data.consensus_meter?.total_papers ?? (papers.length || 1);
-
-          const consensusData = {
-            meter,
-            totalStudies,
-            verified: true,
-            paperUrl
-          };
-
-          finding.consensus = consensusData;
-          finding.qa_verified = true;
-          if (paperUrl) {
-            finding.doi_url = paperUrl;
-          }
-
-          await setCachedRecord(storage, cacheKey, {
-            query: cleanQuery || claimText,
-            consensus: consensusData,
-            cached_at: new Date().toISOString()
-          });
-        }
-      } catch (err) {
-        // Fall back gracefully on network error or offline mode
-      }
-    }
-  }
-
-  return findings;
-}
-
-module.exports = {
-  verifyFindingsWithConsensus,
-  hashClaim,
-  detectNeuromyth,
-  autoCorrectNeuromyth,
-  formatConsensus
-};
diff --git a/extension/shared/dossier-compiler.cjs b/extension/shared/dossier-compiler.cjs
deleted file mode 100644
index 1ded244..0000000
--- a/extension/shared/dossier-compiler.cjs
+++ /dev/null
@@ -1,96 +0,0 @@
-function compileSingleAuditToMarkdown(item) {
-  if (!item || !item.result) return '';
-  const result = item.result;
-  const title = item.title || 'Course Material';
-  const pageType = item.pageType || 'Web Page';
-  const url = item.url || '';
-  const timestamp = (item.timestamp ? new Date(item.timestamp) : new Date()).toLocaleString();
-
-  let md = `# idstack Instructional Design Audit: ${title}\n\n`;
-  md += `> **Audited Component:** ${pageType}  \n`;
-  if (url) md += `> **Source URL:** [${url}](${url})  \n`;
-  md += `> **Date:** ${timestamp}  \n`;
-  md += `> **Evaluator:** idstack Evidence-Based Course Design Engine (Manifest V3)\n\n`;
-
-  md += `## Executive Summary\n\n`;
-  md += `- **Bloom's Demand:** ${result.summary?.bloomsLevel || 'N/A'}\n`;
-  md += `- **Constructive Alignment:** ${result.summary?.alignmentScore || 'N/A'}\n`;
-  md += `- **Key Takeaway:** ${result.summary?.keyTakeaway || 'No summary provided.'}\n\n`;
-
-  if (Array.isArray(result.findings) && result.findings.length > 0) {
-    md += `## Evidence-Based Findings & Recommendations\n\n`;
-    result.findings.forEach((f, idx) => {
-      md += `### ${idx + 1}. [${f.tier || 'T1'}] ${f.citation || 'Citation'}\n\n`;
-      md += `- **Severity:** \`${(f.severity || 'info').toUpperCase()}\`\n`;
-      md += `- **Observation:** ${f.observation || ''}\n`;
-      md += `- **Empirical Evidence:** ${f.evidence || ''}\n`;
-      md += `- **Actionable Recommendation:** ${f.recommendation || ''}\n\n`;
-    });
-  }
-
-  if (result.improvedDraft && result.improvedDraft.content) {
-    md += `## ${result.improvedDraft.title || 'Improved Draft & Alignment Matrix'}\n\n`;
-    md += `${result.improvedDraft.content}\n\n`;
-  }
-
-  md += `---\n*Generated by [idstack](https://github.com/savvides/idstack) — Evidence-Based Course Design.*`;
-  return md;
-}
-
-function compileDossierToMarkdown(dossierItems, courseTitle = 'Canvas Course') {
-  if (!Array.isArray(dossierItems) || dossierItems.length === 0) {
-    return `# idstack Course Audit Dossier: ${courseTitle}\n\n*No audit materials in dossier.*`;
-  }
-
-  const timestamp = new Date().toLocaleString();
-  let md = `# idstack Course Audit Dossier: ${courseTitle}\n\n`;
-  md += `> **Course:** ${courseTitle}  \n`;
-  md += `> **Total Audited Materials:** ${dossierItems.length} components  \n`;
-  md += `> **Compiled Date:** ${timestamp}  \n`;
-  md += `> **Engine:** idstack Evidence-Based Instructional Design Co-Pilot\n\n`;
-
-  md += `## Table of Audited Materials\n\n`;
-  md += `| # | Component Title | Type | Bloom's Demand | Alignment Score |\n`;
-  md += `| :--- | :--- | :--- | :--- | :--- |\n`;
-  dossierItems.forEach((item, idx) => {
-    const s = item.result?.summary || {};
-    md += `| ${idx + 1} | ${item.title || 'Untitled'} | ${item.pageType || 'Page'} | ${s.bloomsLevel || 'N/A'} | ${s.alignmentScore || 'N/A'} |\n`;
-  });
-  md += `\n---\n\n`;
-
-  dossierItems.forEach((item, idx) => {
-    md += `## Section ${idx + 1}: ${item.title || 'Component'}\n\n`;
-    md += `> **Type:** ${item.pageType || 'Page'} | **Source:** ${item.url || 'N/A'}\n\n`;
-
-    const r = item.result || {};
-    md += `### Summary\n`;
-    md += `- **Cognitive Demand:** ${r.summary?.bloomsLevel || 'N/A'}\n`;
-    md += `- **Alignment Rating:** ${r.summary?.alignmentScore || 'N/A'}\n`;
-    md += `- **Takeaway:** ${r.summary?.keyTakeaway || 'N/A'}\n\n`;
-
-    if (Array.isArray(r.findings) && r.findings.length > 0) {
-      md += `### Findings & Evidence\n\n`;
-      r.findings.forEach((f, fIdx) => {
-        md += `#### ${idx + 1}.${fIdx + 1} [${f.tier || 'T1'}] ${f.citation || 'Citation'}\n`;
-        md += `- **Observation:** ${f.observation || ''}\n`;
-        md += `- **Evidence:** ${f.evidence || ''}\n`;
-        md += `- **Recommendation:** ${f.recommendation || ''}\n\n`;
-      });
-    }
-
-    if (r.improvedDraft && r.improvedDraft.content) {
-      md += `### ${r.improvedDraft.title || 'Revised Draft'}\n\n`;
-      md += `${r.improvedDraft.content}\n\n`;
-    }
-
-    md += `---\n\n`;
-  });
-
-  md += `*Generated by [idstack](https://github.com/savvides/idstack) — Evidence-Based Course Design.*`;
-  return md;
-}
-
-module.exports = {
-  compileSingleAuditToMarkdown,
-  compileDossierToMarkdown
-};
diff --git a/extension/shared/evidence-base.cjs b/extension/shared/evidence-base.cjs
deleted file mode 100644
index 0bb8374..0000000
--- a/extension/shared/evidence-base.cjs
+++ /dev/null
@@ -1,26 +0,0 @@
-const TIER_METADATA = {
-  T1: { label: 'Meta-analysis', description: 'Systematic reviews / meta-analyses with large effect sizes', color: '#2f7a4a' },
-  T2: { label: 'Controlled trial', description: 'Peer-reviewed empirical randomized or quasi-experimental studies', color: '#2864a8' },
-  T3: { label: 'Observational', description: 'Correlational, cohort, or longitudinal learning studies', color: '#a87726' },
-  T4: { label: 'Case study', description: 'Single-institution or discipline-specific qualitative studies', color: '#b35a1f' },
-  T5: { label: 'Expert guidance', description: 'Established instructional design frameworks (QM, OLC, Bloom)', color: '#6b6b6b' }
-};
-
-const EVIDENCE_DOMAINS = [
-  { code: 'Models', name: 'Instructional Design Models & Frameworks', keyStudies: ['Abuhassna et al. (2024) [T3]', 'Kalonde et al. (2025) [T3]', 'Crompton et al. (2023) [T3]'] },
-  { code: 'Alignment', name: 'Constructive Alignment & Learning Objectives', keyStudies: ['Biggs (1996) [T5]', 'Anderson & Krathwohl (2001) [T5]', 'Agarwal (2019) [T1]'] },
-  { code: 'Needs', name: 'Needs Analysis', keyStudies: ['Markaki et al. (2021) [T3]', 'Garavan et al. (2019) [T3]', 'Alsalamah & Callinan (2021) [T3]'] },
-  { code: 'Cognitive', name: 'Cognitive Load Theory & Sequencing', keyStudies: ['Sweller (1994) [T5]', 'Costley et al. (2023) [T1]', 'Paas & van Merrienboer (2020) [T5]', 'Chen et al. (2018) [T1]'] },
-  { code: 'Assessment', name: 'Formative Assessment & Feedback', keyStudies: ['Wisniewski et al. (2020) [T1]', 'Hattie & Timperley (2007) [T1]', 'Black & Wiliam (1998) [T1]', 'Double et al. (2019) [T1]'] },
-  { code: 'Multimedia', name: 'Multimedia Learning Principles', keyStudies: ['Mayer (2024) [T5]', 'Moreno & Mayer (1999) [T1]', 'Noetel et al. (2021) [T1]'] },
-  { code: 'Learner', name: 'Learner Analysis & Differentiation', keyStudies: ['Deunk et al. (2018) [T1]', 'Puzio et al. (2020) [T1]', 'Liou et al. (2023) [T2]'] },
-  { code: 'Evaluation', name: 'Evaluation Models', keyStudies: ['Kirkpatrick / Alsalamah & Callinan (2021) [T3]', 'Frye & Hemmer (2012) [T5]', 'Allen et al. (2021) [T3]'] },
-  { code: 'Prototype', name: 'Rapid Prototyping & Design-Based Research', keyStudies: ['Tripp & Bichelmeyer (1990) [T5]', 'Shakeel et al. (2022) [T2]', 'Design-Based Research Collective (2003) [T5]'] },
-  { code: 'Online', name: 'Online Course Quality Frameworks', keyStudies: ['Quality Matters / Zimmerman et al. (2020) [T4]', 'Castro & Tumibay (2019) [T1]', 'Swan et al. (2012) [T2]'] },
-  { code: 'Accessibility', name: 'Universal Design for Learning & A11y', keyStudies: ['CAST UDL Guidelines (2024) [T5]', 'WCAG 2.1/2.2 AA [T5]', 'Puzio et al. (2020) [T1]'] }
-];
-
-module.exports = {
-  TIER_METADATA,
-  EVIDENCE_DOMAINS
-};
diff --git a/extension/shared/evidence-base.js b/extension/shared/evidence-base.js
index 6c01133..615a9ae 100644
--- a/extension/shared/evidence-base.js
+++ b/extension/shared/evidence-base.js
@@ -1,21 +1,7 @@
 export const TIER_METADATA = {
-  T1: { label: 'Meta-analysis', description: 'Systematic reviews / meta-analyses with large effect sizes', color: '#2f7a4a' },
-  T2: { label: 'Controlled trial', description: 'Peer-reviewed empirical randomized or quasi-experimental studies', color: '#2864a8' },
-  T3: { label: 'Observational', description: 'Correlational, cohort, or longitudinal learning studies', color: '#a87726' },
-  T4: { label: 'Case study', description: 'Single-institution or discipline-specific qualitative studies', color: '#b35a1f' },
-  T5: { label: 'Expert guidance', description: 'Established instructional design frameworks (QM, OLC, Bloom)', color: '#6b6b6b' }
+  T1: { label: 'Meta-analyses, RCTs', description: 'RCTs, meta-analyses with learning outcome measures', color: '#2f7a4a' },
+  T2: { label: 'Quasi-experimental with controls', description: 'Quasi-experimental with appropriate controls', color: '#2864a8' },
+  T3: { label: 'Systematic reviews of mixed evidence', description: 'Systematic reviews (synthesis of mixed evidence)', color: '#a87726' },
+  T4: { label: 'Observational, no comparison group', description: 'Observational / pre-post without comparison groups', color: '#b35a1f' },
+  T5: { label: 'Expert opinion, theoretical frameworks', description: 'Expert opinion, literature reviews, theoretical frameworks', color: '#6b6b6b' }
 };
-
-export const EVIDENCE_DOMAINS = [
-  { code: 'Models', name: 'Instructional Design Models & Frameworks', keyStudies: ['Abuhassna et al. (2024) [T3]', 'Kalonde et al. (2025) [T3]', 'Crompton et al. (2023) [T3]'] },
-  { code: 'Alignment', name: 'Constructive Alignment & Learning Objectives', keyStudies: ['Biggs (1996) [T5]', 'Anderson & Krathwohl (2001) [T5]', 'Agarwal (2019) [T1]'] },
-  { code: 'Needs', name: 'Needs Analysis', keyStudies: ['Markaki et al. (2021) [T3]', 'Garavan et al. (2019) [T3]', 'Alsalamah & Callinan (2021) [T3]'] },
-  { code: 'Cognitive', name: 'Cognitive Load Theory & Sequencing', keyStudies: ['Sweller (1994) [T5]', 'Costley et al. (2023) [T1]', 'Paas & van Merrienboer (2020) [T5]', 'Chen et al. (2018) [T1]'] },
-  { code: 'Assessment', name: 'Formative Assessment & Feedback', keyStudies: ['Wisniewski et al. (2020) [T1]', 'Hattie & Timperley (2007) [T1]', 'Black & Wiliam (1998) [T1]', 'Double et al. (2019) [T1]'] },
-  { code: 'Multimedia', name: 'Multimedia Learning Principles', keyStudies: ['Mayer (2024) [T5]', 'Moreno & Mayer (1999) [T1]', 'Noetel et al. (2021) [T1]'] },
-  { code: 'Learner', name: 'Learner Analysis & Differentiation', keyStudies: ['Deunk et al. (2018) [T1]', 'Puzio et al. (2020) [T1]', 'Liou et al. (2023) [T2]'] },
-  { code: 'Evaluation', name: 'Evaluation Models', keyStudies: ['Kirkpatrick / Alsalamah & Callinan (2021) [T3]', 'Frye & Hemmer (2012) [T5]', 'Allen et al. (2021) [T3]'] },
-  { code: 'Prototype', name: 'Rapid Prototyping & Design-Based Research', keyStudies: ['Tripp & Bichelmeyer (1990) [T5]', 'Shakeel et al. (2022) [T2]', 'Design-Based Research Collective (2003) [T5]'] },
-  { code: 'Online', name: 'Online Course Quality Frameworks', keyStudies: ['Quality Matters / Zimmerman et al. (2020) [T4]', 'Castro & Tumibay (2019) [T1]', 'Swan et al. (2012) [T2]'] },
-  { code: 'Accessibility', name: 'Universal Design for Learning & A11y', keyStudies: ['CAST UDL Guidelines (2024) [T5]', 'WCAG 2.1/2.2 AA [T5]', 'Puzio et al. (2020) [T1]'] }
-];
diff --git a/extension/shared/prompts.cjs b/extension/shared/prompts.cjs
deleted file mode 100644
index 3192726..0000000
--- a/extension/shared/prompts.cjs
+++ /dev/null
@@ -1,93 +0,0 @@
-const { EVIDENCE_DOMAINS, TIER_METADATA } = require('./evidence-base.cjs');
-
-function buildAuditPrompt({ title, pageType, content = '' }) {
-  return `You are idstack, an evidence-based instructional design co-pilot.
-Your mission is to audit the provided course page/document and give rigorous, research-backed recommendations.
-
-Target Document Title: "${title || 'Untitled Course Page'}"
-Detected Document Type: ${pageType || 'Course Content'}
-
-Document Content:
-"""
-${(content || '').slice(0, 10000)}
-"""
-
-Please audit this material against peer-reviewed instructional design evidence:
-1. Classify learning objectives or implied cognitive depth using Bloom's Revised Taxonomy (Remember, Understand, Apply, Analyze, Evaluate, Create).
-2. Check Constructive Alignment: Do activities and assessments match the stated or necessary cognitive depth?
-3. Flag Cognitive Load, Elaborated Feedback gaps, and Accessibility considerations.
-4. Rate each recommendation with an evidence tier [T1] to [T5].
-5. Provide a ready-to-use, improved version (rewritten rubric, upgraded learning outcome verbs, or enhanced prompt).
-
-You MUST respond strictly with valid JSON conforming to this schema:
-{
-  "summary": {
-    "bloomsLevel": "Remember | Understand | Apply | Analyze | Evaluate | Create",
-    "alignmentScore": "Strong | Moderate | Weak",
-    "keyTakeaway": "1-2 sentence executive summary of findings"
-  },
-  "findings": [
-    {
-      "severity": "critical | warning | suggestion",
-      "tier": "T1 | T2 | T3 | T4 | T5",
-      "citation": "[Domain-Code] Short Citation",
-      "observation": "What is present in the current material",
-      "evidence": "What peer-reviewed research indicates",
-      "recommendation": "Specific actionable suggestion"
-    }
-  ],
-  "improvedDraft": {
-    "title": "Improved Rubric / Learning Objective / Assignment Prompt",
-    "content": "Full markdown text ready for the instructor to copy-paste into Canvas"
-  }
-}`;
-}
-
-function buildCourseAuditPrompt(courseData = {}) {
-  const assignmentsSummary = (courseData.assignments || [])
-    .map((a, i) => `Assignment ${i+1}: ${a.title} (${a.points || 0} pts)\nDescription: ${(a.description || '').slice(0, 500)}`)
-    .join('\n\n');
-
-  return `You are an expert instructional designer and cognitive scientist using the idstack evidence base.
-Perform a full-course Constructive Alignment audit (courseAudit) for the following course:
-
-COURSE TITLE: ${courseData.title || 'Canvas Course'}
-SYLLABUS & LEARNING OBJECTIVES:
-${(courseData.syllabus || 'No syllabus provided').slice(0, 8000)}
-
-COURSE ASSIGNMENTS & ASSESSMENTS (${(courseData.assignments || []).length} items):
-${assignmentsSummary.slice(0, 20000)}
-
-Evaluate whether the assessment system constructively aligns with stated learning outcomes (Biggs 1996 [T2], Liou et al. 2023 [T2]).
-Identify cognitive load bottlenecks (Sweller 2011 [T1]), scaffolding gaps (Wood et al. 1976 [T2]), and formative feedback quality (Wisniewski et al. 2020 [T1]).
-
-Respond with ONLY a valid JSON object matching this schema:
-{
-  "summary": {
-    "bloomsLevel": "Overall Cognitive Demand (e.g. Apply / Analyze)",
-    "alignmentScore": "High (90%) | Moderate (70%) | Low (40%)",
-    "keyTakeaway": "1-2 sentence executive summary of course-wide curriculum alignment."
-  },
-  "findings": [
-    {
-      "severity": "critical" | "warning" | "info",
-      "tier": "T1" | "T2" | "T3" | "T4" | "T5",
-      "citation": "[Domain-ID] Citation Name",
-      "observation": "What was identified across the course syllabus and assignments.",
-      "evidence": "Author (Year) [Tier description]: Empirical finding.",
-      "recommendation": "Concrete actionable curriculum fix."
-    }
-  ],
-  "improvedDraft": {
-    "title": "Course Alignment & Scaffolding Matrix",
-    "content": "Markdown formatted course roadmap and revised assessment scaffolding."
-  }
-}`;
-}
-
-module.exports = {
-  EVIDENCE_DOMAINS,
-  TIER_METADATA,
-  buildAuditPrompt,
-  buildCourseAuditPrompt
-};
diff --git a/extension/shared/prompts.js b/extension/shared/prompts.js
index 9822f9f..8d142e7 100644
--- a/extension/shared/prompts.js
+++ b/extension/shared/prompts.js
@@ -1,4 +1,6 @@
-import { EVIDENCE_DOMAINS, TIER_METADATA } from './evidence-base.js';
+import { TIER_METADATA } from './evidence-base.js';
+
+const TIER_SCALE = Object.entries(TIER_METADATA).map(([tier, m]) => `   ${tier}: ${m.description}`).join('\n');
 
 export function buildAuditPrompt({ title, pageType, content = '' }) {
   return `You are idstack, an evidence-based instructional design co-pilot.
@@ -16,7 +18,8 @@ Please audit this material against peer-reviewed instructional design evidence:
 1. Classify learning objectives or implied cognitive depth using Bloom's Revised Taxonomy (Remember, Understand, Apply, Analyze, Evaluate, Create).
 2. Check Constructive Alignment: Do activities and assessments match the stated or necessary cognitive depth?
 3. Flag Cognitive Load, Elaborated Feedback gaps, and Accessibility considerations.
-4. Rate each recommendation with an evidence tier [T1] to [T5].
+4. Rate each recommendation with an evidence tier:
+${TIER_SCALE}
 5. Provide a ready-to-use, improved version (rewritten rubric, upgraded learning outcome verbs, or enhanced prompt).
 
 You MUST respond strictly with valid JSON conforming to this schema:
@@ -28,7 +31,7 @@ You MUST respond strictly with valid JSON conforming to this schema:
   },
   "findings": [
     {
-      "severity": "critical | warning | suggestion",
+      "severity": "critical | warning | info",
       "tier": "T1 | T2 | T3 | T4 | T5",
       "citation": "[Domain-Code] Short Citation",
       "observation": "What is present in the current material",
@@ -44,9 +47,21 @@ You MUST respond strictly with valid JSON conforming to this schema:
 }
 
 export function buildCourseAuditPrompt(courseData = {}) {
-  const assignmentsSummary = (courseData.assignments || [])
-    .map((a, i) => `Assignment ${i+1}: ${a.title} (${a.points || 0} pts)\nDescription: ${(a.description || '').slice(0, 500)}`)
-    .join('\n\n');
+  const blocks = (courseData.assignments || [])
+    .map((a, i) => `Assignment ${i+1}: ${a.title} (${a.points || 0} pts)\nDescription: ${(a.description || '').slice(0, 500)}`);
+  // Whole assignments up to a 20,000-char budget, and the header says how many
+  // made it: a plain character cut ended mid-assignment under a header that
+  // still counted every item. The budget keeps a 500-assignment course's
+  // prompt, and so the model's response time, bounded.
+  const total = blocks.length;
+  let shown = 0;
+  let assignmentsSummary = '';
+  for (const block of blocks) {
+    const next = shown ? `${assignmentsSummary}\n\n${block}` : block;
+    if (next.length > 20000) break;
+    assignmentsSummary = next;
+    shown++;
+  }
 
   return `You are an expert instructional designer and cognitive scientist using the idstack evidence base.
 Perform a full-course Constructive Alignment audit (courseAudit) for the following course:
@@ -55,11 +70,13 @@ COURSE TITLE: ${courseData.title || 'Canvas Course'}
 SYLLABUS & LEARNING OBJECTIVES:
 ${(courseData.syllabus || 'No syllabus provided').slice(0, 8000)}
 
-COURSE ASSIGNMENTS & ASSESSMENTS (${(courseData.assignments || []).length} items):
-${assignmentsSummary.slice(0, 20000)}
+COURSE ASSIGNMENTS & ASSESSMENTS (${total} items; ${shown} shown in full below):
+${assignmentsSummary}
 
-Evaluate whether the assessment system constructively aligns with stated learning outcomes (Biggs 1996 [T2], Liou et al. 2023 [T2]).
-Identify cognitive load bottlenecks (Sweller 2011 [T1]), scaffolding gaps (Wood et al. 1976 [T2]), and formative feedback quality (Wisniewski et al. 2020 [T1]).
+Evaluate whether the assessment system constructively aligns with stated learning outcomes ([Alignment-1] Biggs (1996) [T5]).
+Identify cognitive load bottlenecks ([CogLoad-4] Sweller (1994) [T5]), scaffolding gaps ([CogLoad-1] Costley et al. (2023) [T1]), and formative feedback quality ([Assessment-8] Wisniewski et al. (2020) [T1]).
+Rate each finding's evidence tier on this scale:
+${TIER_SCALE}
 
 Respond with ONLY a valid JSON object matching this schema:
 {
@@ -85,4 +102,4 @@ Respond with ONLY a valid JSON object matching this schema:
 }`;
 }
 
-export { EVIDENCE_DOMAINS, TIER_METADATA };
+export { TIER_METADATA };
diff --git a/extension/sidepanel/index.html b/extension/sidepanel/index.html
index 0a5a794..4c75369 100644
--- a/extension/sidepanel/index.html
+++ b/extension/sidepanel/index.html
@@ -13,7 +13,7 @@
   <header class="app-header">
     <div class="brand">
       <span class="logo">idstack</span>
-      <span class="badge-version">v1.0</span>
+      <span class="badge-version">v1.1</span>
     </div>
     <div class="header-actions">
       <button id="dossier-toggle-btn" class="dossier-pill" title="Course Audit Dossier">
@@ -98,19 +98,19 @@ <h3>Settings & Privacy</h3>
           <label for="api-key-input">Google AI Studio API Key (Optional BYOK)</label>
           <input type="password" id="api-key-input" placeholder="AIzaSy..." />
           <a id="get-api-key-link" href="https://aistudio.google.com/app/apikey" target="_blank" rel="noopener noreferrer" class="help-link">Get your free Google AI Studio key &rarr;</a>
-          <p class="help-text">Leave blank to use the built-in free demo tier.</p>
+          <p class="help-text">Without a key, idstack runs in demo mode: every audit shows the same sample findings, whatever the page says, and the page's text is not sent anywhere.</p>
         </div>
         <div class="form-group">
           <label for="consensus-api-key-input">Consensus API Key (Optional BYOK)</label>
           <input type="password" id="consensus-api-key-input" placeholder="cons_..." />
           <a id="get-consensus-api-key-link" href="https://consensus.app" target="_blank" rel="noopener noreferrer" class="help-link">Get your Consensus API key &rarr;</a>
-          <p class="help-text">Enables live academic literature verification via Consensus search.</p>
+          <p class="help-text">Checks each finding against the research literature on Consensus. With a key saved, each finding's claim is sent to Consensus under your key.</p>
         </div>
         <button id="save-settings-btn" class="secondary-btn">Save Settings</button>
         <hr class="divider" />
         <div class="privacy-note">
-          <h4>Privacy & FERPA Commitment</h4>
-          <p>idstack processes only curriculum and assignment text. No student PII is ever collected, retained, or used for model training.</p>
+          <h4>Privacy & Student Data</h4>
+          <p>With a key saved, each audit sends the page's title and text (for a full-course audit, the syllabus and assignments) from your browser to Google's API. On Google's free tier, Google may use that content to improve its products, and human reviewers may read it. idstack will not read Canvas grades, Gradebook, SpeedGrader, People, groups, discussions, submissions, or Inbox pages. Other pages can still name students, so audit course materials, not student work. With a Consensus key saved, each finding's claim, which can quote the page, is also sent to Consensus.</p>
         </div>
       </div>
     </section>
diff --git a/extension/sidepanel/renderer-helper.cjs b/extension/sidepanel/renderer-helper.cjs
deleted file mode 100644
index 8fefa61..0000000
--- a/extension/sidepanel/renderer-helper.cjs
+++ /dev/null
@@ -1,137 +0,0 @@
-/**
- * idstack Side Panel HTML Renderer Helper (CommonJS)
- */
-
-function escapeHtml(str) {
-  if (str === null || str === undefined) return '';
-  return String(str)
-    .replace(/&/g, '&amp;')
-    .replace(/</g, '&lt;')
-    .replace(/>/g, '&gt;')
-    .replace(/"/g, '&quot;')
-    .replace(/'/g, '&#039;');
-}
-
-function renderAuditHTML(data) {
-  if (!data) return '';
-  const summary = data.summary || { bloomsLevel: 'N/A', alignmentScore: 'N/A', keyTakeaway: '' };
-  const findings = data.findings || [];
-  const improvedDraft = data.improvedDraft || { title: 'Improved Draft', content: '' };
-
-  const findingsHTML = findings.map(f => {
-    const tier = escapeHtml(f.tier || 'T1');
-    const tierClass = escapeHtml((f.tier || 'T1').toLowerCase());
-    const severity = escapeHtml(f.severity || 'info');
-    const citation = escapeHtml(f.citation || '');
-    const observation = escapeHtml(f.observation || '');
-    const evidence = escapeHtml(f.evidence || '');
-    const recommendation = escapeHtml(f.recommendation || '');
-
-    let consensusBadgeHTML = '';
-    if (f.consensus) {
-      let badgeText = '✓ Consensus Verified';
-      if (f.consensus.meter !== undefined && f.consensus.meter !== null && f.consensus.totalStudies !== undefined && f.consensus.totalStudies !== null) {
-        const count = f.consensus.totalStudies;
-        const countLabel = count === 1 ? '1 paper' : `${count} papers`;
-        badgeText = `✓ ${f.consensus.meter}% Consensus (${countLabel})`;
-      } else if (f.consensus.meter !== undefined && f.consensus.meter !== null) {
-        badgeText = `✓ ${f.consensus.meter}% Consensus`;
-      }
-      consensusBadgeHTML = `<span class="consensus-badge">${escapeHtml(badgeText)}</span>`;
-    }
-
-    let citationHTML = `<span class="citation">${citation}</span>`;
-    if (f.consensus && f.consensus.paperUrl) {
-      const rawUrl = String(f.consensus.paperUrl);
-      if (/^https?:\/\//i.test(rawUrl)) {
-        citationHTML = `<a href="${escapeHtml(rawUrl)}" target="_blank" rel="noopener noreferrer" class="citation">${citation || 'Source Paper'}</a>`;
-      }
-    }
-
-    return `
-    <div class="finding-card severity-${severity}">
-      <div class="finding-header">
-        <span class="tier-badge tier-${tierClass}">${tier}</span>
-        ${consensusBadgeHTML ? `${consensusBadgeHTML}\n        ` : ''}${citationHTML}
-      </div>
-      <p class="finding-obs"><strong>Observation:</strong> ${observation}</p>
-      <p class="finding-evi"><strong>Evidence:</strong> ${evidence}</p>
-      <p class="finding-rec"><strong>Recommendation:</strong> ${recommendation}</p>
-    </div>
-  `;
-  }).join('');
-
-  const bloomsLevel = escapeHtml(summary.bloomsLevel || 'N/A');
-  const alignmentScore = escapeHtml(summary.alignmentScore || 'N/A');
-  const keyTakeaway = escapeHtml(summary.keyTakeaway || '');
-  const draftTitle = escapeHtml(improvedDraft.title || 'Improved Draft');
-  const draftContent = escapeHtml(improvedDraft.content || '');
-
-  return `
-    <div class="context-card summary-card">
-      <div class="score-row">
-        <span class="chip">Bloom's: <strong>${bloomsLevel}</strong></span>
-        <span class="chip">Alignment: <strong>${alignmentScore}</strong></span>
-      </div>
-      <p class="key-takeaway">${keyTakeaway}</p>
-    </div>
-
-    <h3 class="section-title">Evidence-Based Findings (${findings.length})</h3>
-    <div class="findings-list">${findingsHTML}</div>
-
-    <div class="improved-box">
-      <div class="improved-header">
-        <h4>${draftTitle}</h4>
-        <button id="copy-improved-btn" class="secondary-btn">📋 Copy to Clipboard</button>
-      </div>
-      <pre class="improved-content"><code>${draftContent}</code></pre>
-    </div>
-
-    <div class="feedback-row">
-      <span>Was this audit helpful?</span>
-      <button class="feedback-btn" data-vote="up">👍</button>
-      <button class="feedback-btn" data-vote="down">👎</button>
-      <button id="re-audit-btn" class="link-btn">Audit Another Page</button>
-    </div>
-  `;
-}
-
-function renderDossierListHTML(dossierItems) {
-  if (!Array.isArray(dossierItems) || dossierItems.length === 0) {
-    return `
-      <div class="dossier-empty">
-        <p>No items in dossier yet.</p>
-        <p class="help-text">Audit pages and click &ldquo;Add to Dossier&rdquo; to build your course dossier.</p>
-      </div>
-    `;
-  }
-
-  return dossierItems.map((item) => {
-    const id = escapeHtml(item.id || '');
-    const title = escapeHtml(item.title || 'Untitled Material');
-    const pageType = escapeHtml(item.pageType || 'Page');
-    const blooms = escapeHtml(item.result?.summary?.bloomsLevel || 'N/A');
-    const alignment = escapeHtml(item.result?.summary?.alignmentScore || 'N/A');
-
-    return `
-      <div class="dossier-item" data-dossier-id="${id}">
-        <div class="dossier-item-header">
-          <span class="chip">${pageType}</span>
-          <button class="dossier-delete-btn icon-btn" data-dossier-id="${id}" title="Remove from Dossier">✕</button>
-        </div>
-        <h4 class="dossier-item-title">${title}</h4>
-        <div class="dossier-item-meta">
-          <span>Bloom's: <strong>${blooms}</strong></span>
-          <span>Alignment: <strong>${alignment}</strong></span>
-        </div>
-      </div>
-    `;
-  }).join('');
-}
-
-module.exports = {
-  escapeHtml,
-  renderAuditHTML,
-  renderDossierListHTML
-};
-
diff --git a/extension/sidepanel/renderer-helper.js b/extension/sidepanel/renderer-helper.js
index 65ee399..1cca9e6 100644
--- a/extension/sidepanel/renderer-helper.js
+++ b/extension/sidepanel/renderer-helper.js
@@ -15,12 +15,13 @@ export function escapeHtml(str) {
 export function renderAuditHTML(data) {
   if (!data) return '';
   const summary = data.summary || { bloomsLevel: 'N/A', alignmentScore: 'N/A', keyTakeaway: '' };
-  const findings = data.findings || [];
+  const findings = Array.isArray(data.findings) ? data.findings.filter((f) => f && typeof f === 'object') : [];
   const improvedDraft = data.improvedDraft || { title: 'Improved Draft', content: '' };
 
   const findingsHTML = findings.map(f => {
-    const tier = escapeHtml(f.tier || 'T1');
-    const tierClass = escapeHtml((f.tier || 'T1').toLowerCase());
+    const tierLabel = String(f.tier || 'T1');
+    const tier = escapeHtml(tierLabel);
+    const tierClass = escapeHtml(tierLabel.toLowerCase());
     const severity = escapeHtml(f.severity || 'info');
     const citation = escapeHtml(f.citation || '');
     const observation = escapeHtml(f.observation || '');
diff --git a/extension/sidepanel/sidepanel.css b/extension/sidepanel/sidepanel.css
index 2eed18e..c1e0753 100644
--- a/extension/sidepanel/sidepanel.css
+++ b/extension/sidepanel/sidepanel.css
@@ -29,13 +29,12 @@
   /* Severity palette */
   --severity-critical: #8c2515;
   --severity-warning: #7a4f0c;
-  --severity-suggestion: #2f7a4a;
   --severity-info: #344566;
 
   --sev-critical-bg: #f5dcd5;
   --sev-warning-bg: #fbf0d9;
-  --sev-suggestion-bg: #eef6f0;
   --sev-info-bg: #e7eaf1;
+  --consensus-bg: #eef6f0;
 
   /* Typography */
   --font-display: 'Source Serif 4', 'Source Serif Pro', 'Charter', Georgia, serif;
@@ -71,13 +70,12 @@
 
     --severity-critical: #e89786;
     --severity-warning: #e3b25c;
-    --severity-suggestion: #5fc185;
     --severity-info: #9bafd2;
 
     --sev-critical-bg: #2e1612;
     --sev-warning-bg: #2e2210;
-    --sev-suggestion-bg: #14281b;
     --sev-info-bg: #1a2030;
+    --consensus-bg: #14281b;
   }
 }
 
@@ -289,7 +287,7 @@ a.citation:hover {
   font-size: 11px;
   font-weight: 500;
   color: var(--tier-t1);
-  background: var(--sev-suggestion-bg);
+  background: var(--consensus-bg);
   border: 1px solid var(--tier-t1);
   padding: 1px 6px;
   border-radius: 2px;
@@ -310,10 +308,6 @@ a.citation:hover {
   border-left-color: var(--severity-warning);
 }
 
-.finding-card.severity-suggestion {
-  border-left-color: var(--severity-suggestion);
-}
-
 .finding-card.severity-info {
   border-left-color: var(--severity-info);
 }
diff --git a/extension/sidepanel/sidepanel.js b/extension/sidepanel/sidepanel.js
index 38992de..960d046 100644
--- a/extension/sidepanel/sidepanel.js
+++ b/extension/sidepanel/sidepanel.js
@@ -7,6 +7,9 @@ let activePayload = null;
 let activeCourseContext = null;
 let activeAuditResult = null;
 let activeAuditItem = null;
+let lastAuditBtn = null;
+let lastCourseTarget = null; // `${origin}/courses/${id}` of the last course audit; null after a page audit
+let refreshSeq = 0;
 
 function sanitizeFilename(name) {
   return String(name || 'material')
@@ -27,6 +30,19 @@ function downloadMarkdownFile(filename, content) {
   URL.revokeObjectURL(url);
 }
 
+// A missing navigator.clipboard throws inside the try, so it is reported too.
+async function copyWithFeedback(btn, text, idleLabel) {
+  try {
+    await navigator.clipboard.writeText(text);
+    btn.textContent = '✓ Copied!';
+  } catch (e) {
+    btn.textContent = 'Copy failed';
+  }
+  setTimeout(() => {
+    btn.textContent = idleLabel;
+  }, 2000);
+}
+
 export async function updateDossierBadge() {
   const countEl = document.getElementById('dossier-count');
   if (!countEl) return;
@@ -63,10 +79,12 @@ export async function updateDossierUI() {
 
 export async function refreshActiveTab() {
   if (typeof chrome === 'undefined' || !chrome.tabs || !chrome.tabs.query) return;
+  // Tab events can fire back to back; only the newest refresh may write state.
+  const seq = ++refreshSeq;
 
   try {
     const [tab] = await chrome.tabs.query({ active: true, currentWindow: true });
-    if (!tab || !tab.id) return;
+    if (seq !== refreshSeq || !tab || !tab.id) return;
 
     const url = tab.url || '';
     activeCourseContext = detectCourseContext(url);
@@ -75,8 +93,17 @@ export async function refreshActiveTab() {
       auditCourseBtn.style.display = activeCourseContext.isCourseRoot ? 'block' : 'none';
     }
 
+    // No declarative content script: inject the extractor on demand. Its final
+    // statement evaluates to the payload, which comes back as the injection result.
+    // Needs access to the tab: activeTab (clicking the idstack icon on this tab) or a
+    // host permission (*.instructure.com, or a site granted for a course audit).
     try {
-      const response = await chrome.tabs.sendMessage(tab.id, { action: 'EXTRACT_CONTENT' });
+      const [injection] = await chrome.scripting.executeScript({
+        target: { tabId: tab.id },
+        files: ['content/extractor.js']
+      });
+      if (seq !== refreshSeq) return;
+      const response = injection && injection.result;
       if (response) {
         activePayload = response;
         const pageTypeTag = document.getElementById('page-type-tag');
@@ -86,38 +113,23 @@ export async function refreshActiveTab() {
         return;
       }
     } catch (err) {
-      // Content script may not be injected on pre-existing tabs. Attempt programmatic injection.
-      if (chrome.scripting && chrome.scripting.executeScript) {
-        try {
-          await chrome.scripting.executeScript({
-            target: { tabId: tab.id },
-            files: ['content/extractor.js']
-          });
-          const retryResponse = await chrome.tabs.sendMessage(tab.id, { action: 'EXTRACT_CONTENT' });
-          if (retryResponse) {
-            activePayload = retryResponse;
-            const pageTypeTag = document.getElementById('page-type-tag');
-            const pageTitle = document.getElementById('page-title');
-            if (pageTypeTag) pageTypeTag.textContent = activeCourseContext.isCourseRoot ? 'Canvas Course Root' : retryResponse.pageType;
-            if (pageTitle) pageTitle.textContent = retryResponse.title;
-            return;
-          }
-        } catch (injectionErr) {
-          // Tab may be a chrome:// or restricted URL
-        }
-      }
+      // No access to this tab yet (switched to it without clicking the icon) or a
+      // restricted page such as chrome:// or the Chrome Web Store.
     }
 
-    // Fallback payload if content script extraction fails
+    if (seq !== refreshSeq) return;
+    // Fallback payload if extraction fails. Without access Chrome hides tab.title too.
     const pageTypeTag = document.getElementById('page-type-tag');
     const pageTitle = document.getElementById('page-title');
     if (pageTypeTag) pageTypeTag.textContent = activeCourseContext.isCourseRoot ? 'Canvas Course Root' : 'Web Page';
-    if (pageTitle) pageTitle.textContent = tab.title || 'Current Tab';
+    if (pageTitle) pageTitle.textContent = tab.title || 'Click the idstack toolbar icon to read this tab';
     activePayload = {
       url: tab.url || '',
       title: tab.title || 'Current Tab',
       pageType: activeCourseContext.isCourseRoot ? 'Canvas Course Root' : 'Web Page',
-      content: ''
+      content: '',
+      // The service worker shows this instead of auditing an empty page.
+      emptyReason: 'idstack cannot read this tab. If you just switched to it, click the idstack toolbar icon, then audit again. Browser pages such as chrome:// and the Chrome Web Store cannot be read.'
     };
   } catch (e) {
     console.warn('Error refreshing active tab:', e);
@@ -130,13 +142,15 @@ export function showState(stateName) {
   if (target) target.classList.add('active');
 }
 
-export function renderResults(data) {
+// source: the page (or course) the request was sent for. The active tab may
+// have changed while the audit ran.
+export function renderResults(data, source) {
   activeAuditResult = data;
   activeAuditItem = {
-    id: activePayload?.url || `item-${Date.now()}`,
-    title: activePayload?.title || 'Course Material',
-    pageType: activePayload?.pageType || 'Web Page',
-    url: activePayload?.url || '',
+    id: source?.url || `item-${Date.now()}`,
+    title: source?.title || 'Course Material',
+    pageType: source?.pageType || 'Web Page',
+    url: source?.url || '',
     result: data,
     timestamp: new Date().toISOString()
   };
@@ -145,6 +159,9 @@ export function renderResults(data) {
   if (addToDossierBtn) {
     addToDossierBtn.textContent = '➕ Add to Dossier';
   }
+  // renderError hides the bar; '' restores the stylesheet's display.
+  const actionsBar = document.querySelector('.result-actions-bar');
+  if (actionsBar) actionsBar.style.display = '';
 
   const container = document.getElementById('results-container');
   if (!container) return;
@@ -155,13 +172,7 @@ export function renderResults(data) {
   if (copyBtn) {
     copyBtn.addEventListener('click', () => {
       const contentToCopy = (data && data.improvedDraft && data.improvedDraft.content) || '';
-      if (navigator.clipboard && navigator.clipboard.writeText) {
-        navigator.clipboard.writeText(contentToCopy);
-      }
-      copyBtn.textContent = '✓ Copied!';
-      setTimeout(() => {
-        copyBtn.textContent = '📋 Copy to Clipboard';
-      }, 2000);
+      copyWithFeedback(copyBtn, contentToCopy, '📋 Copy to Clipboard');
     });
   }
 
@@ -174,11 +185,13 @@ export function renderResults(data) {
   }
 
   container.querySelectorAll('.feedback-btn').forEach(btn => {
-    btn.addEventListener('click', (e) => {
-      const parent = e.target.parentElement;
-      if (parent) {
-        parent.innerHTML = '<em>Thank you for your feedback!</em>';
-      }
+    btn.addEventListener('click', () => {
+      const row = btn.parentElement;
+      if (!row) return;
+      // Replace only the vote controls; the row also holds #re-audit-btn.
+      row.querySelectorAll('.feedback-btn').forEach((b) => b.remove());
+      const prompt = row.querySelector('span');
+      if (prompt) prompt.innerHTML = '<em>Thank you for your feedback!</em>';
     });
   });
 
@@ -186,6 +199,11 @@ export function renderResults(data) {
 }
 
 export function renderError(errorMessage) {
+  // Clear the last result so Add to Dossier / Export .md cannot act on it.
+  activeAuditItem = null;
+  const actionsBar = document.querySelector('.result-actions-bar');
+  if (actionsBar) actionsBar.style.display = 'none';
+
   const container = document.getElementById('results-container');
   if (!container) return;
 
@@ -216,8 +234,12 @@ export function renderError(errorMessage) {
   if (retryBtn) {
     retryBtn.addEventListener('click', () => {
       showState('ready');
-      const btn = document.getElementById('audit-btn');
-      if (btn) btn.click();
+      // The course button reads activeCourseContext when clicked. After a tab switch to
+      // another course, or off Canvas, re-clicking it would crawl, and ask for access
+      // to, whatever is active now instead of the course that failed.
+      const ctx = activeCourseContext || {};
+      if (lastCourseTarget && lastCourseTarget !== `${ctx.origin}/courses/${ctx.courseId}`) return;
+      if (lastAuditBtn) lastAuditBtn.click();
     });
   }
 
@@ -260,7 +282,11 @@ if (exportSingleMdBtn) {
 const auditBtn = document.getElementById('audit-btn');
 if (auditBtn) {
   auditBtn.addEventListener('click', async () => {
+    lastAuditBtn = auditBtn;
+    lastCourseTarget = null;
     if (!activePayload) await refreshActiveTab();
+    // Label the result with the page sent, not the tab active when the reply lands.
+    const sent = activePayload;
 
     const progressCard = document.getElementById('crawl-progress-card');
     if (progressCard) progressCard.style.display = 'none';
@@ -272,13 +298,13 @@ if (auditBtn) {
 
     if (typeof chrome !== 'undefined' && chrome.runtime && chrome.runtime.sendMessage) {
       try {
-        chrome.runtime.sendMessage({ action: 'RUN_AUDIT', payload: activePayload }, (response) => {
+        chrome.runtime.sendMessage({ action: 'RUN_AUDIT', payload: sent }, (response) => {
           if (chrome.runtime.lastError) {
             renderError(chrome.runtime.lastError.message);
             return;
           }
           if (response && response.success) {
-            renderResults(response.data);
+            renderResults(response.data, sent);
           } else {
             renderError(response?.error || 'Unknown error occurred during audit.');
           }
@@ -294,12 +320,23 @@ if (auditBtn) {
 const auditCourseBtn = document.getElementById('audit-course-btn');
 if (auditCourseBtn) {
   auditCourseBtn.addEventListener('click', async () => {
-    if (!activePayload) await refreshActiveTab();
-
-    const url = (activePayload && activePayload.url) || '';
-    const courseCtx = activeCourseContext || detectCourseContext(url);
-    const origin = courseCtx.origin || (url ? new URL(url).origin : '');
-    const courseId = courseCtx.courseId;
+    lastAuditBtn = auditCourseBtn;
+    // Shown only after refreshActiveTab() found a course root, so the context is set.
+    const { origin, courseId } = activeCourseContext || {};
+    lastCourseTarget = `${origin}/courses/${courseId}`;
+    // permissions.request needs this click's user gesture, so it comes before any other
+    // await. Already-held access (*.instructure.com) resolves true with no prompt; other
+    // Canvas hosts get a one-time Chrome prompt for that site only.
+    let granted = false;
+    try {
+      granted = await chrome.permissions.request({ origins: [`${origin}/*`] });
+    } catch (err) {
+      // Not a requestable origin (optional_host_permissions covers https only)
+    }
+    if (!granted) {
+      renderError(`idstack needs access to ${origin} to audit the whole course.`);
+      return;
+    }
 
     const progressCard = document.getElementById('crawl-progress-card');
     const statusText = document.getElementById('crawl-status-text');
@@ -337,7 +374,12 @@ if (auditCourseBtn) {
             return;
           }
           if (response && response.success) {
-            renderResults(response.data);
+            // Same label as the worker's auditHistory entry for this course.
+            renderResults(response.data, {
+              url: `${origin}/courses/${courseId}`,
+              title: response.courseData?.title || 'Canvas Course',
+              pageType: 'Canvas Course (Full Audit)'
+            });
           } else {
             renderError(response?.error || 'Unknown error occurred during course audit.');
           }
@@ -395,13 +437,7 @@ if (copyDossierMdBtn) {
       ? `Course ${activeCourseContext.courseId}`
       : (activePayload?.title || 'Canvas Course');
     const md = compileDossierToMarkdown(dossier, courseTitle);
-    if (navigator.clipboard && navigator.clipboard.writeText) {
-      await navigator.clipboard.writeText(md);
-    }
-    copyDossierMdBtn.textContent = '✓ Copied!';
-    setTimeout(() => {
-      copyDossierMdBtn.textContent = '📋 Copy Markdown';
-    }, 2000);
+    await copyWithFeedback(copyDossierMdBtn, md, '📋 Copy Markdown');
   });
 }
 
@@ -484,6 +520,13 @@ if (typeof chrome !== 'undefined' && chrome.tabs) {
   }
 }
 
+// The service worker announces an icon click: activeTab now covers the current tab.
+if (typeof chrome !== 'undefined' && chrome.runtime && chrome.runtime.onMessage) {
+  chrome.runtime.onMessage.addListener((msg) => {
+    if (msg && msg.action === 'TAB_ACCESS_GRANTED') refreshActiveTab();
+  });
+}
+
 // Initial tab detection on load
 if (typeof chrome !== 'undefined' && chrome.tabs) {
   refreshActiveTab();
diff --git a/test/extension-harness.mjs b/test/extension-harness.mjs
new file mode 100644
index 0000000..b077b61
--- /dev/null
+++ b/test/extension-harness.mjs
@@ -0,0 +1,316 @@
+// Loads the SHIPPED extension files under Node with stubbed browser globals.
+// Tests import extension/**/*.js directly; never copy a module into test/ or
+// keep a hand-maintained twin (that is how the old .cjs copies worked).
+//
+// Fake-DOM limits: no entity decoding in text, innerText is not layout-aware,
+// no event bubbling, selectors are limited to tag, #id, .a.b, [attr] and
+// [attr="v"] with descendant combinators and comma lists. The parser expects
+// well-formed markup: a close tag that does not match the open element is
+// ignored, `<x />` on a non-void tag stays open, and an attribute value must
+// not contain the other quote character.
+import { readFileSync } from 'node:fs';
+import { createRequire } from 'node:module';
+import { fileURLToPath } from 'node:url';
+import vm from 'node:vm';
+
+const EXT = new URL('../extension/', import.meta.url);
+const require = createRequire(import.meta.url);
+
+// ES modules (service worker, side panel, shared/*): each call re-runs the
+// module's top-level code. Its imports stay cached, which is fine: they read
+// chrome/fetch/document at call time, not at load time.
+let freshCount = 0;
+export function importFresh(relPath) {
+  return import(`${new URL(relPath, EXT).href}?fresh=${++freshCount}`);
+}
+
+// Classic content scripts carry no ESM syntax, so Node loads them as CommonJS
+// (via extractor.js's module.exports guard). A query string does not bust the
+// CommonJS cache, hence require + cache delete. vm.Script first: Chrome runs
+// content scripts as classic scripts, so import/export must fail here rather
+// than be quietly accepted by Node's require(esm).
+export function loadContentScript(relPath) {
+  const file = fileURLToPath(new URL(relPath, EXT));
+  new vm.Script(readFileSync(file, 'utf8'), { filename: file });
+  delete require.cache[file];
+  return require(file);
+}
+
+// Resolves after every already-queued promise callback has run.
+export const flush = () => new Promise((resolve) => setTimeout(resolve, 0));
+
+export const EXTENSION_ID = 'idstackextensionidfortests';
+// What chrome.runtime.onMessage sees for a message from the side panel page.
+const PANEL_SENDER = { id: EXTENSION_ID, url: `chrome-extension://${EXTENSION_ID}/sidepanel/index.html` };
+
+function storageArea(data) {
+  return {
+    // keys: string | string[] | null (everything) | { key: default }.
+    // Callbacks run asynchronously, as in Chrome: two get-then-set sequences
+    // in flight at once interleave here exactly as they do in the browser.
+    get(keys, callback) {
+      const defaults = keys && typeof keys === 'object' && !Array.isArray(keys) ? keys : {};
+      const wanted = keys == null ? Object.keys(data) : typeof keys === 'string' ? [keys] : Array.isArray(keys) ? keys : Object.keys(keys);
+      const out = {};
+      for (const k of wanted) {
+        if (k in data) out[k] = structuredClone(data[k]);
+        else if (k in defaults) out[k] = defaults[k];
+      }
+      const done = Promise.resolve(out);
+      if (callback) done.then(callback);
+      return done;
+    },
+    set(items, callback) {
+      Object.assign(data, structuredClone(items));
+      const done = Promise.resolve();
+      if (callback) done.then(() => callback());
+      return done;
+    }
+  };
+}
+
+// Installs globalThis.chrome. Options:
+//   sync, local          initial chrome.storage contents
+//   tabs                 what chrome.tabs.query resolves to, or a function
+//                        returning it (read on every query, so a test can switch tabs)
+//   onRuntimeMessage     (message) => response for chrome.runtime.sendMessage
+//   onExecuteScript      (injection) => InjectionResult[] for
+//                        chrome.scripting.executeScript; omitted = reject like
+//                        a tab the extension has no access to
+//   onPermissionRequest  (permissions) => boolean for chrome.permissions.request;
+//                        omitted = granted
+// Returns { chrome, storage, listeners, sent, calls, dispatch }.
+// storage.sync/.local are the live backing objects; sent records
+// runtime.sendMessage calls; calls records executeScript injections
+// (executed), permission requests (permissionRequests), sidePanel.open
+// options (opened) and the last setPanelBehavior value (panelBehavior).
+// dispatch(message, sender = PANEL_SENDER) delivers to the registered
+// runtime.onMessage listeners and resolves with the sendResponse value
+// (undefined if no listener keeps the channel open and none responds).
+export function installChrome({ sync = {}, local = {}, tabs = [], onRuntimeMessage, onExecuteScript, onPermissionRequest } = {}) {
+  const storage = { sync: structuredClone(sync), local: structuredClone(local) };
+  const listeners = { message: [], tabActivated: [], tabUpdated: [], actionClicked: [] };
+  const sent = [];
+  const calls = { executed: [], permissionRequests: [], opened: [], panelBehavior: undefined };
+  const chrome = {
+    runtime: {
+      id: EXTENSION_ID,
+      getURL: (p = '') => `chrome-extension://${EXTENSION_ID}/${p}`,
+      lastError: undefined,
+      onMessage: { addListener: (fn) => listeners.message.push(fn) },
+      sendMessage(message, callback) {
+        sent.push(message);
+        const reply = Promise.resolve(onRuntimeMessage ? onRuntimeMessage(message) : undefined);
+        if (!callback) return reply; // MV3 promise form
+        reply.then(callback);
+      }
+    },
+    storage: { sync: storageArea(storage.sync), local: storageArea(storage.local) },
+    tabs: {
+      query: async () => (typeof tabs === 'function' ? tabs() : tabs),
+      onActivated: { addListener: (fn) => listeners.tabActivated.push(fn) },
+      onUpdated: { addListener: (fn) => listeners.tabUpdated.push(fn) }
+    },
+    scripting: {
+      executeScript: async (injection) => {
+        calls.executed.push(injection);
+        if (!onExecuteScript) throw new Error('Cannot access contents of the page.');
+        return onExecuteScript(injection);
+      }
+    },
+    // Recorded before any await, so a test can check that a request was made
+    // synchronously inside a click handler (it needs the user gesture).
+    permissions: {
+      request(permissions) {
+        calls.permissionRequests.push(permissions);
+        return Promise.resolve(onPermissionRequest ? onPermissionRequest(permissions) : true);
+      }
+    },
+    action: { onClicked: { addListener: (fn) => listeners.actionClicked.push(fn) } },
+    sidePanel: {
+      setPanelBehavior: async (behavior) => { calls.panelBehavior = behavior; },
+      open(options) { calls.opened.push(options); return Promise.resolve(); }
+    }
+  };
+  globalThis.chrome = chrome;
+
+  function dispatch(message, sender = PANEL_SENDER) {
+    return new Promise((resolve) => {
+      let responded = false;
+      const sendResponse = (response) => {
+        if (!responded) { responded = true; resolve(response); }
+      };
+      const keepOpen = listeners.message
+        .map((fn) => fn(message, sender, sendResponse))
+        .some((r) => r === true);
+      if (!keepOpen && !responded) resolve(undefined);
+    });
+  }
+  return { chrome, storage, listeners, sent, calls, dispatch };
+}
+
+// Replaces globalThis.fetch; handler(url, init) returns a Response or throws.
+// Returns the array of recorded { url, init } calls.
+const realFetch = globalThis.fetch;
+export function installFetch(handler) {
+  const calls = [];
+  globalThis.fetch = async (url, init = {}) => {
+    calls.push({ url: String(url), init });
+    return handler(String(url), init);
+  };
+  return calls;
+}
+
+export function jsonResponse(body, init = {}) {
+  return new Response(JSON.stringify(body), {
+    status: 200,
+    headers: { 'Content-Type': 'application/json' },
+    ...init
+  });
+}
+
+// ---- Minimal DOM: enough for extractor.js selectors and sidepanel.js. ----
+// Results come in document order, as in Chrome. Assigning innerHTML re-parses,
+// so rendered buttons are live. Text is not entity-decoded: assert on structure
+// (querySelector) rather than on decoded text.
+const VOID_TAGS = new Set(['meta', 'link', 'img', 'br', 'hr', 'input']);
+
+// What click listeners threw or rejected with. Chrome reports these as uncaught
+// errors, so a test asserts this stays empty rather than letting one crash it.
+export const clickErrors = [];
+// A suite that never inspects clickErrors still fails on one.
+process.on('exit', () => {
+  if (clickErrors.length) {
+    console.error('A click handler threw or rejected:', clickErrors);
+    process.exitCode = 1;
+  }
+});
+
+class FakeElement {
+  constructor(tagName, attrs = {}) {
+    this.tagName = tagName.toLowerCase();
+    this.attrs = attrs;
+    this.children = [];
+    this.parentElement = null;
+    this.listeners = {};
+    this.value = attrs.value || '';
+    this.style = {};
+    for (const decl of (attrs.style || '').split(';')) {
+      const [prop, val] = decl.split(':').map((s) => s && s.trim());
+      if (prop && val) this.style[prop] = val;
+    }
+    const el = this;
+    this.classList = {
+      contains: (c) => el._classes().includes(c),
+      add: (c) => { if (!el.classList.contains(c)) el.attrs.class = [...el._classes(), c].join(' '); },
+      remove: (c) => { el.attrs.class = el._classes().filter((x) => x !== c).join(' '); },
+      toggle: (c) => (el.classList.contains(c) ? (el.classList.remove(c), false) : (el.classList.add(c), true))
+    };
+  }
+  _classes() { return (this.attrs.class || '').split(/\s+/).filter(Boolean); }
+  get id() { return this.attrs.id || ''; }
+  getAttribute(name) { return name in this.attrs ? this.attrs[name] : null; }
+  get textContent() {
+    return this.children.map((c) => (typeof c === 'string' ? c : c.textContent)).join(' ').trim();
+  }
+  set textContent(text) { this.children = [String(text)]; }
+  get innerText() { return this.textContent; }
+  set innerHTML(html) {
+    this.children = parseHtml(String(html)).children;
+    for (const c of this.children) if (typeof c !== 'string') c.parentElement = this;
+  }
+  appendChild(child) { child.parentElement = this; this.children.push(child); return child; }
+  removeChild(child) { this.children = this.children.filter((c) => c !== child); return child; }
+  remove() { if (this.parentElement) this.parentElement.removeChild(this); }
+  addEventListener(type, fn) { (this.listeners[type] ||= []).push(fn); }
+  // Each listener runs synchronously up to its first await, as in Chrome. The
+  // returned promise settles once every (async) listener has finished.
+  click() {
+    const event = { type: 'click', target: this, currentTarget: this, stopPropagation() {}, preventDefault() {} };
+    return Promise.all((this.listeners.click || []).map(async (fn) => {
+      try { await fn(event); } catch (e) { clickErrors.push(e); }
+    }));
+  }
+  _descendants() {
+    const out = [];
+    const walk = (el) => { for (const c of el.children) if (typeof c !== 'string') { out.push(c); walk(c); } };
+    walk(this);
+    return out;
+  }
+  _matches(part) {
+    if (part.startsWith('#')) return this.attrs.id === part.slice(1);
+    if (part.startsWith('.')) return part.split('.').filter(Boolean).every((c) => this._classes().includes(c));
+    if (part.startsWith('[')) {
+      const m = part.match(/\[([a-zA-Z0-9_-]+)(?:="([^"]*)")?\]/);
+      return !!m && (m[2] === undefined ? m[1] in this.attrs : this.attrs[m[1]] === m[2]);
+    }
+    return this.tagName === part.toLowerCase();
+  }
+  // Document order, like the real DOM: querySelector('a, b') returns whichever
+  // matching element comes first in the page, not the first selector's match.
+  querySelectorAll(selectorList) {
+    const selectors = selectorList.split(',').map((s) => s.trim().split(/\s+/));
+    return this._descendants().filter((el) => selectors.some((parts) => el._matchesSelector(parts)));
+  }
+  querySelector(selectorList) { return this.querySelectorAll(selectorList)[0] || null; }
+  _matchesSelector(parts) {
+    if (!this._matches(parts[parts.length - 1])) return false;
+    let i = parts.length - 2;
+    for (let anc = this.parentElement; anc && i >= 0; anc = anc.parentElement) {
+      if (anc._matches(parts[i])) i--;
+    }
+    return i < 0;
+  }
+}
+
+function parseHtml(html) {
+  const root = new FakeElement('#root');
+  const stack = [root];
+  const tagRe = /<(\/)?([a-zA-Z0-9]+)([^>]*)>|([^<]+)/g;
+  let m;
+  while ((m = tagRe.exec(html)) !== null) {
+    const [, isClose, tag, attrStr, text] = m;
+    const top = stack[stack.length - 1];
+    if (text) {
+      if (text.trim()) top.children.push(text.trim());
+    } else if (isClose) {
+      if (stack.length > 1 && top.tagName === tag.toLowerCase()) stack.pop();
+    } else {
+      const attrs = {};
+      const attrRe = /([a-zA-Z0-9_-]+)(?:=["']([^"']*)["'])?/g;
+      let a;
+      while ((a = attrRe.exec(attrStr)) !== null) attrs[a[1]] = a[2] || '';
+      const el = new FakeElement(tag, attrs);
+      top.appendChild(el);
+      if (!VOID_TAGS.has(el.tagName)) stack.push(el);
+    }
+  }
+  return root;
+}
+
+// Builds a document from an HTML string (does not touch globals).
+export function createDocument(html) {
+  const root = parseHtml(html);
+  const titleEl = root.querySelector('title');
+  return {
+    title: titleEl ? titleEl.textContent : '',
+    body: root.querySelector('body') || root,
+    getElementById: (id) => root._descendants().find((el) => el.id === id) || null,
+    querySelector: (s) => root.querySelector(s),
+    querySelectorAll: (s) => root.querySelectorAll(s),
+    createElement: (tag) => new FakeElement(tag)
+  };
+}
+
+// Installs globalThis.document; defaults to the shipped side panel page.
+export function installDocument(html = readFileSync(new URL('sidepanel/index.html', EXT), 'utf8')) {
+  globalThis.document = createDocument(html);
+  return globalThis.document;
+}
+
+export function resetGlobals() {
+  delete globalThis.chrome;
+  delete globalThis.document;
+  delete globalThis.window;
+  globalThis.fetch = realFetch;
+}
diff --git a/test/mutation-test.sh b/test/mutation-test.sh
index 18489b3..935ae4e 100755
--- a/test/mutation-test.sh
+++ b/test/mutation-test.sh
@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# Mutation test: reintroduce each bug fixed in v3.3.0.0 into a throwaway copy
+# Mutation test: reintroduce each fixed defect into a throwaway copy
 # of the repo and confirm the guarding test FAILS. A test that still passes
 # with its bug reintroduced is not guarding anything — that is exactly how the
 # v3.3.0.0 defects survived (the version-classifier suite tested a local copy
@@ -19,6 +19,16 @@ trap 'rm -rf "$WORK"' EXIT
 pass=0; fail=0; skip=0
 expect_fail() { # <name> <cmd...>
   local name="$1"; shift
+  # A mutation that edits nothing proves nothing, whatever the suite then does.
+  # Cases 16 and 17 went stale twice this way: their anchor stopped matching,
+  # python's assert exited nonzero, nothing checked it, and both kept reporting
+  # GUARDED. *.bak is excluded because sed -i.bak writes one even on no match.
+  local d=0
+  diff -rq -x '*.bak' "$WORK/base" "$WORK/r" >/dev/null 2>&1 || d=$?
+  if [ "$d" -ne 1 ]; then
+    echo "  ERROR: $name: the mutation did not change the copy (diff exit $d; stale anchor?) — fix the case; no result is trustworthy until then" >&2
+    exit 1
+  fi
   if "$@" >/dev/null 2>&1; then
     echo "  NOT-GUARDED: $name (test still passed with the bug reintroduced)"
     fail=$((fail+1))
@@ -39,16 +49,20 @@ skip_case() { # <name> <why>
 # 3.12+. Mutating it under 3.12 would report a false NOT-GUARDED.
 PY_LT_312=$(python3 -c 'import sys; print(1 if sys.version_info < (3,12) else 0)' 2>/dev/null || echo 0)
 
+# Every case starts from one pristine snapshot, taken once. It is the tree the
+# baseline below vouches for and the tree expect_fail diffs each mutation
+# against, so all three agree by construction even if $SRC is edited mid-run.
+# Copy only what the suites need; skip .git and any nested worktrees.
+mkdir -p "$WORK/base"
+for item in bin skills templates test evidence docs extension .claude-plugin \
+            VERSION CHANGELOG.md README.md PRIVACY.md TODOS.md CONTRIBUTING.md \
+            DESIGN.md ROADMAP.md CLAUDE.md setup; do
+  [ -e "$SRC/$item" ] && cp -R "$SRC/$item" "$WORK/base/"
+done
+
 fresh() {
   rm -rf "$WORK/r"
-  mkdir -p "$WORK/r"
-  # Copy only what the suites need; skip .git and any nested worktrees.
-  for item in bin skills templates test evidence docs extension .claude-plugin \
-              VERSION CHANGELOG.md README.md TODOS.md CONTRIBUTING.md \
-              DESIGN.md ROADMAP.md CLAUDE.md setup; do
-    [ -e "$SRC/$item" ] && cp -R "$SRC/$item" "$WORK/r/"
-  done
-  return 0
+  cp -R "$WORK/base" "$WORK/r"
 }
 
 # Regenerate after mutating a .tmpl or a spliced template. Without this, the
@@ -67,16 +81,22 @@ regen() {
   fi
 }
 
-# Null-mutation control. expect_fail only checks that smoke-test exits non-zero, so if the
-# UNMUTATED copy already fails, every GUARDED below is vacuous — the suite would report a
-# perfect score while proving nothing. This actually happened: `extension` was missing from
-# fresh()'s copy list, so test-extension.sh failed on every copy and all 36 mutations reported
-# GUARDED regardless of whether their guard worked. Prove the baseline is green before trusting
-# a single result.
+# Baseline: the unmutated copy must pass, or every expect_fail below reports
+# GUARDED for the wrong reason. fresh() once omitted extension/, so smoke-test
+# failed on every copy and each smoke-guarded case "passed" while guarding
+# nothing. smoke-test runs every other guarding suite, so one run covers them.
+fresh
+if ! "$WORK/r/test/smoke-test.sh" "$WORK/r" >"$WORK/baseline.log" 2>&1; then
+  echo "  ERROR: the unmutated copy fails smoke-test; every GUARDED result would be meaningless" >&2
+  grep -A5 'FAIL:' "$WORK/baseline.log" >&2
+  exit 1
+fi
+
+# 0. expect_fail must refuse a mutation that changed nothing. Run in a subshell
+# so its exit is observable; bash does not run the parent's EXIT trap there.
 fresh
-if ! "$WORK/r/test/smoke-test.sh" "$WORK/r" >/dev/null 2>&1; then
-  echo "  ABORT: smoke-test fails on an UNMUTATED copy — every GUARDED below would be vacuous." >&2
-  echo "  Run it by hand on \$WORK/r to see which check fails; fresh() is probably missing a path." >&2
+if ( expect_fail "no-op mutation" false ) >/dev/null 2>&1; then
+  echo "  ERROR: a mutation that changed nothing was counted as GUARDED" >&2
   exit 1
 fi
 
@@ -315,13 +335,11 @@ expect_fail "second-CLI dist/ bundle returns" "$WORK/r/test/smoke-test.sh" "$WOR
 fresh
 python3 - "$WORK/r/docs/index.html" <<'PY'
 import sys
-p = sys.argv[1]; s = open(p).read()
-old = '<h2 id="install-title">Get started in seconds.</h2>'
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = '<h2 id="install-title">'
 assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
-s = s.replace(old,
-              '<h2 id="install-title">Get started in seconds.</h2>\n'
-              "      <p>Also runs in OpenAI Codex CLI.</p>", 1)  # IDSTACK_CLI_LEAK_ALLOW
-open(p,'w').write(s)
+s = s.replace(old, '<p>Also runs in OpenAI Codex CLI.</p>\n        ' + old, 1)  # IDSTACK_CLI_LEAK_ALLOW
+open(p, 'w', encoding='utf-8').write(s)
 PY
 expect_fail "untagged reference in a file with tagged lines" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
@@ -334,13 +352,11 @@ expect_fail "untagged reference in a file with tagged lines" "$WORK/r/test/smoke
 fresh
 python3 - "$WORK/r/docs/index.html" <<'PY'
 import sys
-p = sys.argv[1]; s = open(p).read()
-old = '<h2 id="install-title">Get started in seconds.</h2>'
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = '<h2 id="install-title">'
 assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
-s = s.replace(old,
-              '<h2 id="install-title">Get started in seconds.</h2>\n'
-              "      <p>Reviewed by Gemini Code Assist. Also runs in Codex CLI.</p>", 1)  # IDSTACK_CLI_LEAK_ALLOW
-open(p,'w').write(s)
+s = s.replace(old, '<p>Reviewed by Gemini Code Assist. Also runs in Codex CLI.</p>\n        ' + old, 1)  # IDSTACK_CLI_LEAK_ALLOW
+open(p, 'w', encoding='utf-8').write(s)
 PY
 expect_fail "claim carrying the bot's name is not exempt" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
@@ -417,7 +433,964 @@ open(p, 'w').write(s)
 PY
 expect_fail "dateModified reintroduced into docs/index.html" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 23. The notch-safe gutter drops back to a plain var(--pad-x), leaving .section
+# 23. an untagged claim in a committed doc under docs/superpowers/ -> smoke-test
+# must fail. 2b36bea added --exclude-dir=superpowers to the sweep to hide the
+# extension's design docs; grep matches that name at any depth, so it exempted
+# every committed spec and plan under docs/superpowers/ and superpowers/.
+fresh
+printf 'idstack also runs in OpenAI Codex CLI.\n' > "$WORK/r/docs/superpowers/leak.md"  # IDSTACK_CLI_LEAK_ALLOW
+expect_fail "committed docs/superpowers is not sweep-exempt" "$WORK/r/test/smoke-test.sh" "$WORK/r"
+
+# 24a/24b/24f. The extension suite must exercise the SHIPPED files. Every
+# extension test used to load a hand-maintained .cjs twin, so a bug planted in
+# the .js that Chrome actually runs stayed green.
+
+# 24a. shipped renderer stops escaping '<' -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/renderer-helper.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "    .replace(/</g, '&lt;')\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "shipped renderer stops escaping '<'" "$WORK/r/test/test-extension.sh"
+
+# 24b. the course-root detector the side panel imports treats every course page
+# as the root -> test-extension must fail. The old suite tested extractor.js's
+# copy (via a .cjs twin); sidepanel.js imports content/extractor-core.js.
+fresh
+python3 - "$WORK/r/extension/content/extractor-core.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "isCourseRoot: !!match,"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "isCourseRoot: !!anyCourseMatch,", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "side panel course-root detector matches every course page" "$WORK/r/test/test-extension.sh"
+
+# 24f. a hand-copied .cjs twin comes back -> test-extension must fail.
+fresh
+cp "$WORK/r/extension/sidepanel/renderer-helper.js" "$WORK/r/extension/sidepanel/renderer-helper.cjs"
+expect_fail "hand-copied .cjs twin back in extension/" "$WORK/r/test/test-extension.sh"
+
+# 24c/24d/24e. service-worker.js, storage.js and sidepanel.js never ran under
+# the old suite (no chrome or DOM stub), so a bug planted in any of them stayed
+# green. Each case now trips a test that executes the shipped module.
+
+# 24c. RUN_AUDIT stops holding the reply channel open -> test-extension must
+# fail. Without `return true` Chrome closes the port and the panel gets undefined.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "      return true; // async reply\n    }\n\n    if (request.action === 'CRAWL_AND_AUDIT_COURSE')"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "    }\n\n    if (request.action === 'CRAWL_AND_AUDIT_COURSE')", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "service worker drops return true for RUN_AUDIT" "$WORK/r/test/test-extension.sh"
+
+# 24d. re-auditing a page appends a duplicate dossier entry -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/shared/storage.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (existingIdx >= 0) {"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "if (existingIdx > 0) {", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "dossier replace-by-url off by one" "$WORK/r/test/test-extension.sh"
+
+# 24e. side panel hides the course-audit button on a course root -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "isCourseRoot ? 'block' : 'none'"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "isCourseRoot ? 'none' : 'block'", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course-audit button hidden on a course root" "$WORK/r/test/test-extension.sh"
+
+# 25a-25d. The full-course audit read one page of 50 assignments and turned any
+# fetch failure into an empty list, and the prompt then cut that list at 20,000
+# characters mid-assignment under a header that still counted every item.
+
+# 25a. the crawler stops following Link rel="next" -> test-crawler must fail.
+# Canvas pages assignments 50 at a time; reading only page 1 audited the first
+# 50 of an 80-assignment course without saying so.
+fresh
+python3 - "$WORK/r/extension/background/canvas-crawler.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "assignmentsUrl = next ? next[1] : null;"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "assignmentsUrl = null;", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "canvas crawler ignores Link rel=next" node "$WORK/r/test/test-crawler.mjs"
+
+# 25b. the original try/catch returns around the assignments fetch -> test-crawler
+# must fail. That catch turned a 403 or a dropped connection into an empty list,
+# and the audit then told the instructor the course had no assessments.
+fresh
+python3 - "$WORK/r/extension/background/canvas-crawler.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+for old, new in [
+    ("  for (let page = 1; assignmentsUrl; page++) {\n",
+     "  try {\n  for (let page = 1; assignmentsUrl; page++) {\n"),
+    ("    assignmentsUrl = next ? next[1] : null;\n  }\n",
+     "    assignmentsUrl = next ? next[1] : null;\n  }\n  } catch (e) {\n    console.warn('Could not fetch assignments list:', e);\n  }\n"),
+]:
+    assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+    s = s.replace(old, new, 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "canvas crawler swallows assignment-fetch failures" node "$WORK/r/test/test-crawler.mjs"
+
+# 25c. pagination page cap removed -> test-crawler must fail. The loop follows
+# server-supplied links; without the cap a Link cycle crawls until OOM. The test
+# fake is finite (12 pages) so this fails in milliseconds instead of hanging.
+fresh
+python3 - "$WORK/r/extension/background/canvas-crawler.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (page > MAX_ASSIGNMENT_PAGES) {"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "if (false) {", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "canvas crawler page cap removed" node "$WORK/r/test/test-crawler.mjs"
+
+# 25d. the course prompt goes back to a plain 20,000-char slice of every
+# assignment -> test-prompts must fail. The slice ended mid-assignment while the
+# header counted every item. Reverting needs both edits: the whole-block budget
+# already fits in 20,000 chars, so the slice alone would change nothing.
+fresh
+python3 - "$WORK/r/extension/shared/prompts.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+for old, new in [
+    ("    if (next.length > 20000) break;\n", ""),
+    ("${assignmentsSummary}\n", "${assignmentsSummary.slice(0, 20000)}\n"),
+]:
+    assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+    s = s.replace(old, new, 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course prompt cuts the assignment list mid-block" node "$WORK/r/test/test-prompts.mjs"
+
+# 26a-26g. The extension shipped a tier scale that put randomized trials in T2,
+# demo findings citing Biggs as [Alignment-3] [T2] and a nonexistent
+# [Cognitive-2], uncoded prompt citations, and a page prompt asking for a
+# 'suggestion' severity. test-evidence-labels.mjs checks every label against
+# evidence/references.md and CLAUDE.md; each case reverts one defect.
+
+# 26a. TIER_METADATA relabels T2 as randomized trials -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/shared/evidence-base.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "T2: { label: 'Quasi-experimental with controls', description: 'Quasi-experimental with appropriate controls'"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "T2: { label: 'Controlled trial', description: 'Peer-reviewed empirical randomized or quasi-experimental studies'", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "tier scale puts randomized trials in T2" "$WORK/r/test/test-extension.sh"
+
+# 26b. the demo cites Biggs as [Alignment-3] [T2] again -> test-extension must
+# fail. references.md files Biggs as [Alignment-1] T5; [Alignment-3] is Pereira, T4.
+fresh
+python3 - "$WORK/r/extension/background/parser-helper.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "tier: 'T5',\n        citation: '[Alignment-1] Direct Constructive Alignment',\n        observation: 'Learning objectives"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "tier: 'T2',\n        citation: '[Alignment-3] Direct Constructive Alignment',\n        observation: 'Learning objectives", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "demo cites Biggs as [Alignment-3] T2" "$WORK/r/test/test-extension.sh"
+
+# 26c. the demo cites the nonexistent [Cognitive-2] again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/background/parser-helper.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "'[CogLoad-1] Cognitive Load & Chunking'"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "'[Cognitive-2] Cognitive Load & Chunking'", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "demo cites nonexistent [Cognitive-2]" "$WORK/r/test/test-extension.sh"
+
+# 26d. the course prompt cites Biggs uncoded at T2 again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/shared/prompts.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "([Alignment-1] Biggs (1996) [T5])."
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "(Biggs 1996 [T2], Liou et al. 2023 [T2]).", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course prompt cites Biggs uncoded at T2" "$WORK/r/test/test-extension.sh"
+
+# 26e. the page prompt asks for a 'suggestion' severity again -> test-extension
+# must fail. The course prompt, the reports and the side panel use critical|warning|info.
+fresh
+python3 - "$WORK/r/extension/shared/prompts.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = '"severity": "critical | warning | info",'
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, '"severity": "critical | warning | suggestion",', 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "page prompt asks for a suggestion severity" "$WORK/r/test/test-extension.sh"
+
+# 26f. the course prompt drops the canonical tier scale -> test-extension must
+# fail. Without it the model rates tiers on whatever scale it assumes.
+fresh
+python3 - "$WORK/r/extension/shared/prompts.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "Rate each finding's evidence tier on this scale:\n${TIER_SCALE}\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course prompt drops the tier scale" "$WORK/r/test/test-extension.sh"
+
+# 26g. the side panel styles a 'suggestion' finding card again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.css" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = ".finding-card.severity-info {"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, ".finding-card.severity-suggestion {\n  border-left-color: #2f7a4a;\n}\n\n.finding-card.severity-info {", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "side panel styles a suggestion severity" "$WORK/r/test/test-extension.sh"
+
+# 27a-27f. The service worker audited whatever it was sent: a restricted tab's
+# empty fallback, a lone Modules item title or an empty course came back as an
+# invented audit and was saved. Malformed model JSON was saved and stuck the
+# panel, and the model fetch had no timeout and carried the API key in its URL.
+# test-service-worker.mjs drives the shipped worker through each case.
+
+# 27a. RUN_AUDIT audits a page with no readable text again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (wordCount < MIN_AUDIT_WORDS) {"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "if (false) {", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "RUN_AUDIT audits a page with no text" "$WORK/r/test/test-extension.sh"
+
+# 27b. malformed model JSON is saved and handed to the panel again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "throw new Error('The AI response did not match the audit format. Please retry the audit.');"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "malformed model JSON saved to history" "$WORK/r/test/test-extension.sh"
+
+# 27c. the API key goes back into the URL query string -> test-extension must
+# fail. A query-string key lands in proxy, gateway and crash logs.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+for old, new in [
+    (":generateContent';", ":generateContent?key=' + apiKey;"),
+    (", 'x-goog-api-key': apiKey", ""),
+]:
+    assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+    s = s.replace(old, new, 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "API key sent in the URL query string" "$WORK/r/test/test-extension.sh"
+
+# 27d. the model fetch loses its timeout -> test-extension must fail. Chrome
+# kills a service worker whose fetch response takes over 30 s, and the panel
+# then gets a closed message port instead of an explanation.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "const signal = AbortSignal.timeout(LLM_TIMEOUT_MS);"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "const signal = new AbortController().signal;", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "model fetch without a timeout" "$WORK/r/test/test-extension.sh"
+
+# 27e. the worker drops the reason the extractor or panel gave for an unread
+# page -> test-extension must fail. The user would see a generic word count.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "request.payload?.emptyReason || "
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "refusal reason from the extractor/panel dropped" "$WORK/r/test/test-extension.sh"
+
+# 27f. a course with no syllabus and no assignments is audited again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (!courseData.syllabus.trim() && courseData.assignments.length === 0) {"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "if (false) {", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "empty course audited" "$WORK/r/test/test-extension.sh"
+
+# 28a-28n. The extension declared an <all_urls> content script (install prompt:
+# "all your data on all websites"), yet custom-domain Canvas hosts could not be
+# crawled; and the service worker answered any sender, splicing the raw origin
+# and courseId into credentialed Canvas fetches. Now the panel reads tabs through
+# activeTab (the icon click reaches action.onClicked), requests a Canvas host per
+# site, and the worker serves only its own pages and validated course targets.
+
+# 28a. the worker gates on sender.id, which a content script shares -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (!String(sender.url || '').startsWith(chrome.runtime.getURL(''))) return;"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "if (sender.id !== chrome.runtime.id) return;", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "service worker trusts content-script senders" "$WORK/r/test/test-extension.sh"
+
+# 28b. the course origin may carry a path or query again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "target.origin !== origin || "
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course origin smuggles a path" "$WORK/r/test/test-extension.sh"
+
+# 28c. the course id may carry a path again -> test-extension must fail. fetch
+# normalizes '../' segments, so '../../users/self' is any credentialed GET on the host.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = " || !/^\\d+$/.test(courseId)"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course id smuggles a path" "$WORK/r/test/test-extension.sh"
+
+# 28d. the <all_urls> content script comes back -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/manifest.json" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = '  "permissions": [\n'
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, '  "content_scripts": [\n    {\n      "matches": ["<all_urls>"],\n      "js": ["content/extractor.js"],\n      "run_at": "document_idle"\n    }\n  ],\n' + old, 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "<all_urls> content script back" "$WORK/r/test/test-extension.sh"
+
+# 28e. the per-site optional host permission is dropped -> test-extension must
+# fail. Canvas outside instructure.com could then never be crawled.
+fresh
+python3 - "$WORK/r/extension/manifest.json" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = '  ],\n  "optional_host_permissions": [\n    "https://*/*"\n  ]\n'
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, '  ]\n', 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "custom-domain permission dropped" "$WORK/r/test/test-extension.sh"
+
+# 28f. the icon click toggles the panel directly again -> test-extension must
+# fail. Chrome's side-panel toggle path skips the activeTab grant.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "setPanelBehavior({ openPanelOnActionClick: false })"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "setPanelBehavior({ openPanelOnActionClick: true })", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "icon click no longer grants activeTab" "$WORK/r/test/test-extension.sh"
+
+# 28g. the site permission is requested after an await -> test-extension must
+# fail. Chrome rejects permissions.request once the click's user gesture is gone.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "    let granted = false;\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "    await refreshActiveTab();\n" + old, 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "site permission requested after an await" "$WORK/r/test/test-extension.sh"
+
+# 28h. the extractor's final statement no longer yields the payload -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "} else {\n  extractPageContent();\n}\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "}\n", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "extractor no longer returns its payload" "$WORK/r/test/test-extension.sh"
+
+# 28i. a top-level const in the extractor -> test-extension must fail. The panel
+# re-injects into the same isolated world, where a redeclaration throws.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "function detectPageType("
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "const EXTRACTOR_READY = true;\n" + old, 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "extractor breaks on re-injection" "$WORK/r/test/test-extension.sh"
+
+# 28j. an unreadable tab is sent with no explanation -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "      emptyReason: 'idstack cannot read this tab. If you just switched to it, click the idstack toolbar icon, then audit again. Browser pages such as chrome:// and the Chrome Web Store cannot be read.'\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "unreadable tab sent with no explanation" "$WORK/r/test/test-extension.sh"
+
+# 28k. the Chrome floor for sidePanel.open is dropped -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/manifest.json" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = '  "minimum_chrome_version": "116",\n'
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, '', 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "Chrome floor for sidePanel.open dropped" "$WORK/r/test/test-extension.sh"
+
+# 28l. the course crawl ignores a denied site permission -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "    if (!granted) {"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "    if (false) {", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course crawl runs after the site permission is denied" "$WORK/r/test/test-extension.sh"
+
+# 28m. the icon click no longer opens the panel -> test-extension must fail.
+# With openPanelOnActionClick off, onClicked is the only way the panel opens.
+fresh
+python3 - "$WORK/r/extension/background/service-worker.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "  chrome.sidePanel.open({ windowId: tab.windowId });\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "icon click no longer opens the panel" "$WORK/r/test/test-extension.sh"
+
+# 28n. the panel ignores the worker's TAB_ACCESS_GRANTED -> test-extension must
+# fail. After a tab switch and an icon click the header would stay stale.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "  if (msg && msg.action === 'TAB_ACCESS_GRANTED') refreshActiveTab();\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "panel ignores newly granted tab access" "$WORK/r/test/test-extension.sh"
+
+# 29a-29i. The extractor sent a Modules page's first item title only, read the
+# whole page on any Canvas view it did not recognize (Gradebook, People, Inbox,
+# discussions), treated any /courses/ URL as Canvas, and read Google Docs from
+# the editor DOM, which holds no document text. test-extractor.mjs runs the
+# shipped extractor on each page.
+
+# 29a. a Modules page reads only the first item title -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "Array.from(document.querySelectorAll('.module-item-title'))"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "[document.querySelector('.module-item-title')]", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "Modules page reads only the first item title" "$WORK/r/test/test-extension.sh"
+
+# 29b. a course home in Modules view is not detected -> test-extension must fail.
+# Canvas's modules container is #context_modules, and the course home URL has no /modules.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "'#context_modules, #modules'"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "'#modules'", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course home in Modules view not detected" "$WORK/r/test/test-extension.sh"
+
+# 29c. an unrecognized Canvas view falls back to the whole page -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "content = mainBody ? (mainBody.innerText || mainBody.textContent || '').trim() : '';"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "content = mainBody ? (mainBody.innerText || mainBody.textContent || '').trim() : (document.body ? (document.body.innerText || document.body.textContent || '').trim() : '');", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "Canvas page falls back to the whole document body" "$WORK/r/test/test-extension.sh"
+
+# 29d. the student-record guard is disabled -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (studentRecordUrl.test(url)) {"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "if (false && studentRecordUrl.test(url)) {", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "student-record page guard disabled" "$WORK/r/test/test-extension.sh"
+
+# 29e. custom-domain Canvas course pages are not detected as Canvas -> test-extension
+# must fail. They would go through the generic reader, which reads [role=main].
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = r" || /\/courses\/\d+(?:[/?#]|$)/.test(url)"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "custom-domain Canvas course page not detected as Canvas" "$WORK/r/test/test-extension.sh"
+
+# 29f. any /courses/<slug> site is treated as Canvas again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = r"/\/courses\/\d+(?:[/?#]|$)/.test(url)"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "url.includes('/courses/')", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "non-Canvas /courses/<slug> site treated as Canvas" "$WORK/r/test/test-extension.sh"
+
+# 29g. rubric pages lose their text without the body fallback -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = ", #rubrics'"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "'", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "rubric text lost with the body fallback" "$WORK/r/test/test-extension.sh"
+
+# 29h. Google Docs are read from the editor DOM again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "  if (!doc) return data;\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "  return data;\n", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "Google Docs read from the canvas-rendered DOM" "$WORK/r/test/test-extension.sh"
+
+# 29i. the Docs export accepts any content type -> test-extension must fail. A
+# signed-out export answers 200 with an HTML sign-in page, audited as the document.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = " || !(res.headers.get('content-type') || '').startsWith('text/plain')"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "HTML sign-in page audited as the Google Doc" "$WORK/r/test/test-extension.sh"
+
+# 30a-30l. The side panel labelled a result with whatever tab was active when
+# the reply landed, let a slow tab refresh overwrite a newer one, left the last
+# result behind Add to Dossier after an error, retried the single-page audit
+# whatever had failed, wiped Audit Another Page on a vote, and said "Copied!"
+# when the clipboard write failed. The renderer threw on malformed findings,
+# inside the sendMessage callback, so the spinner stayed up.
+# test-sidepanel-state.mjs drives the shipped sidepanel.js through each case.
+
+# 30a. a page audit is labelled with the tab active when the reply lands ->
+# test-sidepanel-state must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "renderResults(response.data, sent);"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "renderResults(response.data, activePayload);", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "page audit labelled with the tab active when the reply lands" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30b. a course audit is labelled with the current page, not the course the
+# worker crawled -> test-sidepanel-state must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = ("renderResults(response.data, {\n"
+       "              url: `${origin}/courses/${courseId}`,\n"
+       "              title: response.courseData?.title || 'Canvas Course',\n"
+       "              pageType: 'Canvas Course (Full Audit)'\n"
+       "            });")
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "renderResults(response.data, activePayload);", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "course audit label ignores the crawled course" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30c. every refresh passes the sequence check -> test-sidepanel-state must
+# fail. A slow injection for a tab the user already left overwrites the newer
+# tab's title and payload, and the next audit sends the wrong page.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "const seq = ++refreshSeq;"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "const seq = refreshSeq;", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "stale tab refresh overwrites a newer tab" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30d. an error leaves the previous result actionable -> test-sidepanel-state
+# must fail. Add to Dossier would save the last success under the failed page.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "\n  activeAuditItem = null;\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "\n", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "error leaves the previous result actionable" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30e. the action bar hidden by an error never comes back -> test-sidepanel-state must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "  if (actionsBar) actionsBar.style.display = '';\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "action bar never returns after an error" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30f. Copy to Clipboard says "Copied!" whether or not the write worked ->
+# test-sidepanel-state must fail. The .catch keeps the failure on the label.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "copyWithFeedback(copyBtn, contentToCopy, '📋 Copy to Clipboard');"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "navigator.clipboard.writeText(contentToCopy).catch(() => {}); copyBtn.textContent = '✓ Copied!';", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "copy button claims success on a rejected write" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30g. Copy Markdown lets a rejected write escape the click handler ->
+# test-sidepanel-state must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "await copyWithFeedback(copyDossierMdBtn, md, '📋 Copy Markdown');"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "await navigator.clipboard.writeText(md); copyDossierMdBtn.textContent = '✓ Copied!';", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "dossier copy leaks an unhandled rejection" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30h. Retry always runs the single-page audit -> test-sidepanel-state must
+# fail. A failed course audit was retried as an audit of the current page.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (lastAuditBtn) lastAuditBtn.click();"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "document.getElementById('audit-btn').click();", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "Retry always runs the single-page audit" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30m. Retry re-clicks the course button after a tab switch -> test-sidepanel-state
+# must fail. The button reads the active course when clicked, so Retry would crawl,
+# and ask for access to, the course or site now active. Both switch scenarios (to a
+# non-course tab, and to another course root, which keeps the button visible) fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "      if (lastCourseTarget && lastCourseTarget !== `${ctx.origin}/courses/${ctx.courseId}`) return;\n"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "Retry audits the course or site now active, not the one that failed" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30i. a feedback vote rewrites the whole row -> test-sidepanel-state must fail.
+# The row also holds Audit Another Page, the only way back to the ready state.
+fresh
+python3 - "$WORK/r/extension/sidepanel/sidepanel.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "if (prompt) prompt.innerHTML = '<em>Thank you for your feedback!</em>';"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "row.innerHTML = '<em>Thank you for your feedback!</em>';", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "feedback vote wipes Audit Another Page" node "$WORK/r/test/test-sidepanel-state.mjs"
+
+# 30j. the renderer trusts a non-array findings value -> test-renderer-helper must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/renderer-helper.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "Array.isArray(data.findings) ? data.findings.filter((f) => f && typeof f === 'object') : []"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "data.findings || []", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "renderer trusts non-array findings" node "$WORK/r/test/test-renderer-helper.mjs"
+
+# 30k. the renderer keeps null and non-object findings -> test-renderer-helper must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/renderer-helper.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = ".filter((f) => f && typeof f === 'object')"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "renderer crashes on a null finding entry" node "$WORK/r/test/test-renderer-helper.mjs"
+
+# 30l. the renderer calls toLowerCase on a non-string tier -> test-renderer-helper must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/renderer-helper.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "String(f.tier || 'T1')"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "(f.tier || 'T1')", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "renderer calls toLowerCase on a non-string tier" node "$WORK/r/test/test-renderer-helper.mjs"
+
+# 31a-31i. The docs made promises the code did not keep. The side panel said
+# no student data was "ever collected, retained, or used for model training"
+# while the extractor read whole Canvas pages and the panel pointed at a free
+# API tier whose terms allow that use. The settings called canned demo output
+# a "free demo tier". PRIVACY.md put the API key in chrome.storage.local and
+# called the design FERPA-compliant. README said the course crawler reads
+# modules, discussions and quizzes. test-disclosures.mjs checks each doc claim
+# against the shipped code.
+
+# 31a. the side panel's absolute no-PII promise returns -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/index.html" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "<h4>Privacy & Student Data</h4>"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, old + "\n          <p>No student PII is ever collected, retained, or used for model training.</p>", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "absolute no-PII promise back in the side panel" "$WORK/r/test/test-extension.sh"
+
+# 31b. the free-tier data-use disclosure is dropped from the side panel -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/index.html" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = " On Google's free tier, Google may use that content to improve its products, and human reviewers may read it."
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "free-tier data-use disclosure dropped from the side panel" "$WORK/r/test/test-extension.sh"
+
+# 31c. demo mode advertised as a service tier again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/extension/sidepanel/index.html" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "Without a key, idstack runs in demo mode: every audit shows the same sample findings, whatever the page says, and the page's text is not sent anywhere."
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "Leave blank to use the built-in free demo tier.", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "demo advertised as a service tier again" "$WORK/r/test/test-extension.sh"
+
+# 31d. the privacy policy names the wrong storage area for the API key -> test-extension must fail.
+fresh
+python3 - "$WORK/r/PRIVACY.md" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "Your Google API key and your Consensus key are saved in `chrome.storage.sync`."
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "Your Google API key and your Consensus key are saved in `chrome.storage.local`.", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "privacy policy names the wrong storage area for the API key" "$WORK/r/test/test-extension.sh"
+
+# 31e. the never-accesses-PII claim is back in the privacy policy -> test-extension must fail.
+fresh
+python3 - "$WORK/r/PRIVACY.md" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "- **No tracking.**"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "- **Curriculum-Only Processing:** The extension only reads public or instructor-accessible course materials (syllabi, module structures, assignment guidelines, and rubrics). It never accesses student rosters, student submissions, student grades, or any Personally Identifiable Information (PII).\n" + old, 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "never-accesses-PII claim back in the privacy policy" "$WORK/r/test/test-extension.sh"
+
+# 31f. audit-history retention goes undisclosed -> test-extension must fail.
+fresh
+python3 - "$WORK/r/PRIVACY.md" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "The last 20 audit results (page address, title, page type, and the full audit result) are kept in `chrome.storage.local`, as is the Course Dossier you build."
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "The Course Dossier you build is kept in `chrome.storage.local`.", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "audit-history retention undisclosed" "$WORK/r/test/test-extension.sh"
+
+# 31g. the README crawler bullet overclaims again -> test-extension must fail.
+fresh
+python3 - "$WORK/r/README.md" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "Reads the course syllabus and assignment descriptions through Canvas's API using your active browser session."
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "Crawls published syllabus items, modules, assignments, discussions, and quizzes in the background using your active browser session.", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "README crawler bullet overclaims again" "$WORK/r/test/test-extension.sh"
+
+# 31h. the docs say Inbox is refused but the extractor reads it -> test-disclosures
+# must fail. Run directly: test-extractor.mjs would also fail under test-extension.sh,
+# leaving GUARDED ambiguous about which assertion caught it.
+fresh
+python3 - "$WORK/r/extension/content/extractor.js" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "|conversations"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "docs say Inbox is refused but the extractor reads it" node "$WORK/r/test/test-disclosures.mjs"
+
+# 31i. the FERPA-compliance claim is back in the privacy policy -> test-extension must fail.
+# idstack cannot certify FERPA compliance; that determination is the institution's.
+fresh
+python3 - "$WORK/r/PRIVACY.md" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "The extension is built to audit course materials, not student records."
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "The idstack Chrome Extension is designed with a strict privacy-first and FERPA-compliant architecture.", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "FERPA-compliance claim back in the privacy policy" "$WORK/r/test/test-extension.sh"
+
+# 31j. the policy stops naming a host the manifest can reach -> test-extension must fail.
+# #89 added api.consensus.app to host_permissions without a word in PRIVACY.md. The
+# plugin section names the host too, so this proves the check reads the extension section.
+fresh
+python3 - "$WORK/r/PRIVACY.md" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = "from your browser to Consensus (`api.consensus.app`)"
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, "from your browser to Consensus", 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "privacy policy omits a host the manifest can reach" "$WORK/r/test/test-extension.sh"
+
+# 32. the side panel's version badge drifts from the manifest -> test-extension must fail.
+# The badge is hand-written, so a manifest bump alone would leave the panel showing the old version.
+fresh
+python3 - "$WORK/r/extension/sidepanel/index.html" <<'PY'
+import sys
+p = sys.argv[1]; s = open(p, encoding='utf-8').read()
+old = '<span class="badge-version">v1.1</span>'
+assert s.count(old) == 1, 'anchor not unique: %d' % s.count(old)
+s = s.replace(old, '<span class="badge-version">v1.0</span>', 1)
+open(p, 'w', encoding='utf-8').write(s)
+PY
+expect_fail "side panel version badge drifts from the manifest" "$WORK/r/test/test-extension.sh"
+
+# 33. The notch-safe gutter drops back to a plain var(--pad-x), leaving .section
 # (which wraps every content region) under the notch once viewport-fit=cover is
 # set. This is the defect the shared container rule fixed.
 fresh
@@ -441,7 +1414,7 @@ open(p, 'w').write(s.replace(old, new, 1))
 PY
 expect_fail "notch-safe gutter loses its env() inset" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 24. A later :root override ships 30px touch targets below 480px while the 44px
+# 34. A later :root override ships 30px touch targets below 480px while the 44px
 # declaration is still present, so an existence check on --tap-min would pass.
 # Proves the assertion requires exactly one declaration, not merely the string.
 fresh
@@ -455,7 +1428,7 @@ open(p, 'w').write(s)
 PY
 expect_fail "a later override shrinks the 44px touch target" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 25. The viewport clip returns. It hides horizontal overflow instead of
+# 35. The viewport clip returns. It hides horizontal overflow instead of
 # preventing it, so a regression becomes unreachable content rather than a
 # visible bug -- measured in Chrome as 305px of the install command unreachable
 # when an element does overflow.
@@ -470,7 +1443,7 @@ open(p, 'w').write(s)
 PY
 expect_fail "root overflow-x clip masks page overflow" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 26. The 44px floor is hoisted out of the unconditional cascade into a
+# 36. The 44px floor is hoisted out of the unconditional cascade into a
 # min-width query, so phones get no floor at all while the token and the
 # declaration both still exist. This is the original scoping bug in mirror
 # image, and an assertion that scans the whole sheet cannot see it.
@@ -487,9 +1460,9 @@ open(p, 'w').write(s)
 PY
 expect_fail "touch-target floor hoisted into a desktop-only query" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 27. A second <style> block overrides the token. Every assertion reads the
+# 37. A second <style> block overrides the token. Every assertion reads the
 # stylesheet, so a scanner that stops at the first block is blind to it --
-# and that blindness would also defeat mutation 24.
+# and that blindness would also defeat mutation 34.
 fresh
 python3 - "$WORK/r/docs/index.html" <<'PY'
 import sys
@@ -501,7 +1474,7 @@ open(p, 'w').write(s)
 PY
 expect_fail "a second <style> block overrides the touch-target token" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 28. The breakpoint rule moves into `@media not all and (max-width: 480px)`,
+# 38. The breakpoint rule moves into `@media not all and (max-width: 480px)`,
 # which applies ABOVE 480px. The prelude still contains the breakpoint string,
 # so a substring match counts it as scoped when it is inverted.
 fresh
@@ -519,7 +1492,7 @@ open(p, 'w').write(s)
 PY
 expect_fail "an inverted @media prelude passes as breakpoint-scoped" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 29. A later base rule at equal specificity re-declares the touch-target floor
+# 39. A later base rule at equal specificity re-declares the touch-target floor
 # in px. The token is untouched and the first .copy-btn block still reads
 # var(--tap-min), so a positive check that stops at the first match cannot see
 # it -- only counting the declarations sheet-wide can.
@@ -534,7 +1507,7 @@ open(p, 'w').write(s)
 PY
 expect_fail "a later base rule outranks the touch-target token" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 30. The viewport clip returns as the `overflow` shorthand rather than the
+# 40. The viewport clip returns as the `overflow` shorthand rather than the
 # `overflow-x` longhand. Same rendered effect, different spelling.
 fresh
 python3 - "$WORK/r/docs/index.html" <<'PY'
@@ -547,7 +1520,7 @@ open(p, 'w').write(s)
 PY
 expect_fail "root overflow shorthand masks page overflow" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 31. The mobile breakpoint stops collapsing to one column. `1fr 1fr` still
+# 41. The mobile breakpoint stops collapsing to one column. `1fr 1fr` still
 # contains `1fr`, so an unanchored match treats the two-column regression as
 # satisfying the single-column assertion.
 fresh
@@ -561,7 +1534,7 @@ open(p, 'w').write(s)
 PY
 expect_fail "pipeline stays two-column at the mobile breakpoint" "$WORK/r/test/smoke-test.sh" "$WORK/r"
 
-# 32. A selector LIST shrinks the touch target below 480px. The text suite anchors every
+# 42. A selector LIST shrinks the touch target below 480px. The text suite anchors every
 # touch-target guard on the single-selector spelling `.copy-btn {`, so `.copy-btn, .btn-badge {`
 # is invisible to all three of its checks at once and it exits 0. Only the rendered suite
 # catches this, by measuring the button at 375px. This mutation is the reason
diff --git a/test/smoke-test.sh b/test/smoke-test.sh
index 3766198..ada06d2 100755
--- a/test/smoke-test.sh
+++ b/test/smoke-test.sh
@@ -323,7 +323,6 @@ fi
 # them a node-less run just reports a smaller total than CLAUDE.md documents, with nothing saying why.
 if command -v node &>/dev/null; then
   check "chrome extension tests pass" "'$IDSTACK_DIR/test/test-extension.sh'"
-  check "consensus extension client tests pass" "node '$IDSTACK_DIR/test/test-consensus-extension.js'"
   check "responsive landing page tests pass" "node '$IDSTACK_DIR/test/test-responsive-landing.js'"
   # Renders the page in headless Chrome and asserts the OUTCOME (no sideways scroll, touch
   # targets, column counts) rather than the CSS text. Skips loudly without a browser; the
@@ -331,7 +330,6 @@ if command -v node &>/dev/null; then
   check "rendered landing page tests pass" "node '$IDSTACK_DIR/test/test-rendered-landing.js'"
 else
   echo "  SKIP: chrome extension tests (node not installed)"
-  echo "  SKIP: consensus extension client tests (node not installed)"
   echo "  SKIP: responsive landing page tests (node not installed)"
   echo "  SKIP: rendered landing page tests (node not installed)"
 fi
@@ -434,21 +432,26 @@ check "preamble embeds no ~/.agents fallbacks" "! grep -q '\.agents/' '$IDSTACK_
 #      directories. Implementation reports and review notes written during
 #      development are not shipped and CI never sees them (fresh checkout), but
 #      they can discuss the sweep and trip it unreliably.
-#   3. Lines tagged IDSTACK_CLI_LEAK_ALLOW. Three kinds of line carry the tag:
-#      this block's own patterns, the dated release note on the landing page,
-#      and the comments crediting "Gemini Code Assist" — a PR-review bot  # IDSTACK_CLI_LEAK_ALLOW
+#      --exclude-dir matches a name at any depth, so never add the dotless
+#      `superpowers`: it exempted the committed specs and plans under
+#      docs/superpowers/ and superpowers/.
+#   3. Lines tagged IDSTACK_CLI_LEAK_ALLOW. Kinds of line that carry the tag
+#      include: this block's own patterns, the dated release note on the
+#      landing page, the comments crediting "Gemini Code Assist" — a PR-review bot  # IDSTACK_CLI_LEAK_ALLOW
 #      that flagged the version classifier four times, unrelated to the CLI and
-#      the reason those test cases exist.
+#      the reason those test cases exist — and mentions of the Gemini model API  # IDSTACK_CLI_LEAK_ALLOW
+#      the Chrome extension calls, which is an LLM endpoint, not a CLI.
 #
-#      A tag is for a dated, historical mention. It is NEVER for a line that
-#      claims idstack runs somewhere it does not. Tag individual lines, never
-#      whole files, and never filter on a bare string: an earlier draft dropped
-#      every line containing the bot's name repo-wide, which would have
-#      let an untagged capability claim through anywhere it appeared.
+#      A tag is for a dated, historical mention or the extension's model API.
+#      It is NEVER for a line that claims idstack runs somewhere it does not.
+#      Tag individual lines, never whole files, and never filter on a bare
+#      string: an earlier draft dropped every line containing the bot's name
+#      repo-wide, which would have let an untagged capability claim through
+#      anywhere it appeared.
 CLI_LEAK_RE='codex|gemini'            # IDSTACK_CLI_LEAK_ALLOW
 CLI_LEAK="$(grep -rIiE "$CLI_LEAK_RE" "$IDSTACK_DIR" \
   --exclude-dir=.git --exclude-dir=.gstack --exclude-dir=.idstack \
-  --exclude-dir=.claude --exclude-dir=.superpowers --exclude-dir=superpowers \
+  --exclude-dir=.claude --exclude-dir=.superpowers \
   --exclude-dir=docs/designs --exclude-dir=designs --exclude=CHANGELOG.md 2>/dev/null || true)"
 CLI_LEAK="$(printf '%s' "$CLI_LEAK" | grep -vF 'IDSTACK_CLI_LEAK_ALLOW' || true)"
 # Printed through the command itself, not tested with -z, so a failure names
diff --git a/test/test-consensus-extension.js b/test/test-consensus-client.mjs
similarity index 99%
rename from test/test-consensus-extension.js
rename to test/test-consensus-client.mjs
index 38b758f..d2fbea7 100644
--- a/test/test-consensus-extension.js
+++ b/test/test-consensus-client.mjs
@@ -1,10 +1,10 @@
-const assert = require('assert');
-const {
+import assert from 'node:assert';
+import {
   verifyFindingsWithConsensus,
   hashClaim,
   detectNeuromyth,
   autoCorrectNeuromyth
-} = require('../extension/shared/consensus-client.cjs');
+} from '../extension/shared/consensus-client.js';
 
 async function runTests() {
   console.log('==> Starting Consensus Extension Client unit tests...');
diff --git a/test/test-course-context.mjs b/test/test-course-context.mjs
new file mode 100644
index 0000000..8c9418a
--- /dev/null
+++ b/test/test-course-context.mjs
@@ -0,0 +1,37 @@
+import assert from 'node:assert';
+import { detectCourseContext } from '../extension/content/extractor-core.js';
+
+// The side panel imports this module (sidepanel.js L3); test that file, not the content script.
+const rootCtx = detectCourseContext('https://canvas.instructure.com/courses/987654', 'Biology 101');
+assert.strictEqual(rootCtx.isCourseRoot, true);
+assert.strictEqual(rootCtx.courseId, '987654');
+assert.strictEqual(rootCtx.origin, 'https://canvas.instructure.com');
+
+const modulesCtx = detectCourseContext('https://canvas.instructure.com/courses/987654/modules', 'Modules');
+assert.strictEqual(modulesCtx.isCourseRoot, true);
+assert.strictEqual(modulesCtx.courseId, '987654');
+assert.strictEqual(modulesCtx.origin, 'https://canvas.instructure.com');
+
+const queryParamCtx = detectCourseContext('https://canvas.instructure.com/courses/987654/modules?view=feed', 'Modules Feed');
+assert.strictEqual(queryParamCtx.isCourseRoot, true);
+assert.strictEqual(queryParamCtx.courseId, '987654');
+
+const subpageCtx = detectCourseContext('https://canvas.instructure.com/courses/987654/assignments/123', 'Lab 1');
+assert.strictEqual(subpageCtx.isCourseRoot, false);
+assert.strictEqual(subpageCtx.courseId, '987654');
+
+// Edge cases from closed bot PR #100.
+for (const bad of [null, 'invalid-url']) {
+  assert.deepStrictEqual(detectCourseContext(bad), { isCourseRoot: false, courseId: null, origin: null }, `${bad} is not a course`);
+}
+const hashCtx = detectCourseContext('https://canvas.instructure.com/courses/987654#main');
+assert.strictEqual(hashCtx.isCourseRoot, true, 'a fragment does not change the page');
+assert.strictEqual(hashCtx.courseId, '987654');
+assert.deepStrictEqual(detectCourseContext('https://example.com/other/123'), { isCourseRoot: false, courseId: null, origin: 'https://example.com' });
+for (const url of ['https://canvas.instructure.com/courses/', 'https://canvas.instructure.com/courses/abc']) {
+  const ctx = detectCourseContext(url);
+  assert.strictEqual(ctx.isCourseRoot, false, `${url} has no numeric course id`);
+  assert.strictEqual(ctx.courseId, null);
+}
+
+console.log('✅ Course context detection tests passed.');
diff --git a/test/test-crawler.js b/test/test-crawler.js
deleted file mode 100644
index 5665e0d..0000000
--- a/test/test-crawler.js
+++ /dev/null
@@ -1,86 +0,0 @@
-const assert = require('assert');
-const { crawlCanvasCourse, stripHtml } = require('../extension/background/canvas-crawler.cjs');
-const { getDemoCourseAuditResult } = require('../extension/background/parser-helper.cjs');
-
-// Test 1: HTML Tag Stripping
-const clean = stripHtml('<p>Hello <strong>World</strong> &amp; Students<br></p>');
-assert.strictEqual(clean, 'Hello World & Students');
-
-// Test 2: Mock Canvas Crawler
-let fetchOptionsPassed = [];
-const mockFetch = async (url, options) => {
-  if (options) fetchOptionsPassed.push(options);
-  if (url.includes('include[]=syllabus_body')) {
-    return {
-      ok: true,
-      json: async () => ({
-        name: 'Biology 101: Cell Systems',
-        syllabus_body: '<p>Welcome to Biology 101. Objectives: Analyze cellular metabolism.</p>'
-      })
-    };
-  }
-  if (url.includes('/assignments')) {
-    return {
-      ok: true,
-      json: async () => [
-        { name: 'Quiz 1', description: '<p>Recall organelles</p>', points_possible: 10 },
-        { name: 'Lab Report 1', description: '<p>Analyze enzyme kinetics</p>', points_possible: 50 }
-      ]
-    };
-  }
-  throw new Error('Not found: ' + url);
-};
-
-(async () => {
-  const courseData = await crawlCanvasCourse('https://canvas.instructure.com', '12345', mockFetch);
-  assert.strictEqual(courseData.title, 'Biology 101: Cell Systems');
-  assert.ok(fetchOptionsPassed.every(opt => opt && opt.credentials === 'include'), 'fetchImpl must include credentials');
-  assert.ok(courseData.syllabus.includes('Analyze cellular metabolism'));
-  assert.strictEqual(courseData.assignments.length, 2);
-  assert.strictEqual(courseData.assignments[0].title, 'Quiz 1');
-  assert.strictEqual(courseData.assignments[0].description, 'Recall organelles');
-
-  // Test 2.5: Missing Error Path Test for crawlCanvasCourse
-  let caughtError1 = false;
-  try {
-    await crawlCanvasCourse(null, '12345', mockFetch);
-  } catch (err) {
-    caughtError1 = true;
-    assert.strictEqual(err.message, 'Canvas origin and courseId are required for course crawling.');
-  }
-  assert.ok(caughtError1, 'Should throw error when origin is missing');
-
-  let caughtError2 = false;
-  try {
-    await crawlCanvasCourse('https://canvas.instructure.com', null, mockFetch);
-  } catch (err) {
-    caughtError2 = true;
-    assert.strictEqual(err.message, 'Canvas origin and courseId are required for course crawling.');
-  }
-  assert.ok(caughtError2, 'Should throw error when courseId is missing');
-
-  // Test 3: Demo Course Audit Fallback
-  const demoResult = getDemoCourseAuditResult({ title: 'Biology 101: Cell Systems' });
-  assert.ok(demoResult.summary.keyTakeaway.includes('Course-Level Demo'));
-  assert.ok(demoResult.findings.length >= 2);
-  assert.ok(demoResult.improvedDraft.title.includes('Matrix'));
-
-  // Test 4: Error Path Test for course fetch failure
-  const mockFailedFetch = async (url) => {
-    return {
-      ok: false,
-      status: 404
-    };
-  };
-
-  let errorThrown = false;
-  try {
-    await crawlCanvasCourse('https://canvas.instructure.com', 'invalid-id', mockFailedFetch);
-  } catch (err) {
-    errorThrown = true;
-    assert.strictEqual(err.message, 'Failed to fetch Canvas course info (404)');
-  }
-  assert.ok(errorThrown, 'Expected an error to be thrown for failed fetch');
-
-  console.log('✅ Task 2 Canvas crawler tests passed.');
-})();
diff --git a/test/test-crawler.mjs b/test/test-crawler.mjs
new file mode 100644
index 0000000..54abf65
--- /dev/null
+++ b/test/test-crawler.mjs
@@ -0,0 +1,109 @@
+import assert from 'node:assert';
+import { crawlCanvasCourse, stripHtml } from '../extension/background/canvas-crawler.js';
+import { getDemoCourseAuditResult } from '../extension/background/parser-helper.js';
+
+// Fake Canvas that behaves like the real API under session-cookie auth: clean
+// JSON bodies (Canvas stopped prefixing while(1); in Aug 2019, canvas-lms
+// commit 485acb0a05), list endpoints paginated through a Link header, and JSON
+// error bodies on 4xx. The previous fake served one clean page and never
+// failed, which is how a crawler that audited a truncated or empty assignment
+// list passed green.
+const ORIGIN = 'https://canvas.test';
+const COURSE_URL = `${ORIGIN}/api/v1/courses/12345?include[]=syllabus_body`;
+const PAGE1_URL = `${ORIGIN}/api/v1/courses/12345/assignments?per_page=50`;
+const pageUrl = (n) => (n === 1 ? PAGE1_URL : `${ORIGIN}/api/v1/courses/12345/assignments?page=${n}&per_page=50`);
+// Canvas sends rel="current" and rel="first" on every page, rel="next" on all but the last.
+const linkFor = (n, last) => [`<${pageUrl(n)}>; rel="current"`]
+  .concat(n < last ? [`<${pageUrl(n + 1)}>; rel="next"`] : [], [`<${PAGE1_URL}>; rel="first"`])
+  .join(',');
+const assignmentsFrom = (first, count) => Array.from({ length: count }, (_, i) => ({
+  name: `Assignment ${first + i}`, description: '<p>Analyze enzyme kinetics</p>', points_possible: 10
+}));
+const COURSE = { body: { name: 'Biology 101: Cell Systems', syllabus_body: '<p>Welcome to Biology 101. Objectives: Analyze cellular metabolism.</p>' } };
+
+function canvasFetch(routes) {
+  const calls = [];
+  const fetchImpl = async (url, options = {}) => {
+    calls.push({ url, credentials: options.credentials });
+    const route = typeof routes === 'function' ? routes(url) : routes[url];
+    if (!route) throw new TypeError(`Failed to fetch ${url}`);
+    const headers = { 'Content-Type': 'application/json; charset=utf-8' };
+    if (route.link) headers.Link = route.link;
+    return new Response(JSON.stringify(route.body), { status: route.status || 200, headers });
+  };
+  return { fetchImpl, calls };
+}
+
+// Test 1: HTML tag stripping
+assert.strictEqual(stripHtml('<p>Hello <strong>World</strong> &amp; Students<br></p>'), 'Hello World & Students');
+
+// Test 2: an 80-assignment course is read across both pages
+{
+  const { fetchImpl, calls } = canvasFetch({
+    [COURSE_URL]: COURSE,
+    [PAGE1_URL]: { body: assignmentsFrom(1, 50), link: linkFor(1, 2) },
+    [pageUrl(2)]: { body: assignmentsFrom(51, 30), link: linkFor(2, 2) }
+  });
+  const courseData = await crawlCanvasCourse(ORIGIN, '12345', fetchImpl);
+  assert.strictEqual(courseData.title, 'Biology 101: Cell Systems');
+  assert.ok(courseData.syllabus.includes('Analyze cellular metabolism'));
+  assert.ok(calls.every((c) => c.credentials === 'include'), 'every Canvas GET must send session cookies');
+  assert.strictEqual(courseData.assignments.length, 80, 'crawler must follow Link rel="next"');
+  assert.strictEqual(courseData.assignments[0].description, 'Analyze enzyme kinetics');
+  assert.strictEqual(courseData.assignments[79].title, 'Assignment 80');
+}
+
+// Test 3: a failed assignments fetch is fatal, never an empty list the LLM
+// reads as "this course has no assessments"
+{
+  const { fetchImpl } = canvasFetch({
+    [COURSE_URL]: COURSE,
+    [PAGE1_URL]: { status: 403, body: { errors: [{ message: 'user not authorized to perform that action' }] } }
+  });
+  await assert.rejects(crawlCanvasCourse(ORIGIN, '12345', fetchImpl), /Failed to fetch Canvas assignments \(403\)/);
+}
+
+// Test 4: a network failure on page 2 is fatal, not a silent 50-item list
+{
+  const { fetchImpl } = canvasFetch({
+    [COURSE_URL]: COURSE,
+    [PAGE1_URL]: { body: assignmentsFrom(1, 50), link: linkFor(1, 2) }
+  });
+  await assert.rejects(crawlCanvasCourse(ORIGIN, '12345', fetchImpl), /Failed to fetch/);
+}
+
+// Test 5: page cap. The fake course is FINITE (12 pages) so a crawler without
+// the cap returns 600 items and fails here instead of looping until OOM.
+{
+  const { fetchImpl, calls } = canvasFetch((url) => {
+    if (url === COURSE_URL) return COURSE;
+    for (let n = 1; n <= 12; n++) {
+      if (url === pageUrl(n)) return { body: assignmentsFrom((n - 1) * 50 + 1, 50), link: linkFor(n, 12) };
+    }
+    return null;
+  });
+  await assert.rejects(crawlCanvasCourse(ORIGIN, '12345', fetchImpl), /more than 500 assignments/);
+  assert.strictEqual(calls.filter((c) => c.url.includes('/assignments')).length, 10);
+}
+
+// Test 6: missing origin or course id is refused before any request
+{
+  const { fetchImpl, calls } = canvasFetch({});
+  await assert.rejects(crawlCanvasCourse(null, '12345', fetchImpl), /Canvas origin and courseId are required/);
+  await assert.rejects(crawlCanvasCourse(ORIGIN, null, fetchImpl), /Canvas origin and courseId are required/);
+  assert.strictEqual(calls.length, 0);
+}
+
+// Test 7: a failed course fetch is fatal
+{
+  const { fetchImpl } = canvasFetch({ [COURSE_URL]: { status: 404, body: { errors: [{ message: 'The specified resource does not exist.' }] } } });
+  await assert.rejects(crawlCanvasCourse(ORIGIN, '12345', fetchImpl), /Failed to fetch Canvas course info \(404\)/);
+}
+
+// Test 8: Demo Course Audit Fallback
+const demoResult = getDemoCourseAuditResult({ title: 'Biology 101: Cell Systems' });
+assert.ok(demoResult.summary.keyTakeaway.includes('Course-Level Demo'));
+assert.ok(demoResult.findings.length >= 2);
+assert.ok(demoResult.improvedDraft.title.includes('Matrix'));
+
+console.log('✅ Task 2 Canvas crawler tests passed.');
diff --git a/test/test-disclosures.mjs b/test/test-disclosures.mjs
new file mode 100644
index 0000000..7630fd9
--- /dev/null
+++ b/test/test-disclosures.mjs
@@ -0,0 +1,121 @@
+// What the side panel, PRIVACY.md and README.md say about the extension must
+// match the shipped code. v3.5.0.0's panel promised no student data was "ever
+// collected, retained, or used for model training", PRIVACY.md put the API key
+// in chrome.storage.local and called the design FERPA-compliant, the settings
+// called canned demo output a "free demo tier", and README said the course
+// crawler reads modules, discussions and quizzes. Each fact below is taken from
+// the shipped modules (storage areas, demo output, crawler requests, extractor
+// refusals), never restated here.
+import assert from 'node:assert';
+import { readFileSync } from 'node:fs';
+import { installChrome, loadContentScript, resetGlobals } from './extension-harness.mjs';
+import { saveSettings, saveAuditResult, addToDossier } from '../extension/shared/storage.js';
+import { getDemoAuditResult, getDemoCourseAuditResult } from '../extension/background/parser-helper.js';
+import { crawlCanvasCourse } from '../extension/background/canvas-crawler.js';
+
+const read = (rel) => readFileSync(new URL(`../${rel}`, import.meta.url), 'utf8');
+const html = read('extension/sidepanel/index.html');
+const privacy = read('PRIVACY.md');
+const readme = read('README.md');
+const manifest = JSON.parse(read('extension/manifest.json'));
+const DOCS = [['index.html', html], ['PRIVACY.md', privacy], ['README.md', readme]];
+
+// Collect every disagreement so one run names them all.
+const problems = new Set();
+const check = (ok, msg) => { if (!ok) problems.add(msg); };
+
+// 1. Storage: record which chrome.storage area each key really lands in, and
+//    how many audits the history keeps.
+const h = installChrome();
+await saveSettings({ apiKey: 'k' });
+for (let i = 0; i < 25; i++) await saveAuditResult({ url: `u${i}`, result: {} });
+await addToDossier({ url: 'u', result: {} });
+resetGlobals();
+const areaOf = (key) => Object.keys(h.storage).find((a) => key in h.storage[a]);
+for (const [key, label] of [['apiKey', /API key/i], ['auditHistory', /audit history/i], ['activeDossier', /dossier/i]]) {
+  const lines = privacy.split('\n').filter((l) => label.test(l) && l.includes('chrome.storage.'));
+  check(lines.length, `PRIVACY.md must say where ${key} is stored (code: chrome.storage.${areaOf(key)})`);
+  for (const l of lines) check(l.includes(`chrome.storage.${areaOf(key)}`), `PRIVACY.md misstates where ${key} is stored (code: chrome.storage.${areaOf(key)}): ${l}`);
+}
+const kept = h.storage[areaOf('auditHistory')].auditHistory.length;
+check(privacy.includes(`last ${kept} audit results`), `PRIVACY.md must disclose that the last ${kept} audit results are kept`);
+
+// 2. Demo mode ignores the page. The docs must say so, and stop saying so if that changes.
+const same = (f, a, b) => JSON.stringify(f(a)) === JSON.stringify(f(b));
+const canned = same(getDemoAuditResult, { title: 'T', pageType: 'P', content: 'alpha' }, { title: 'T', pageType: 'P', content: 'omega' })
+  && same(getDemoCourseAuditResult, { title: 'C', syllabus: 'a', assignments: [] }, { title: 'C', syllabus: 'b', assignments: [{ title: 'x' }] });
+for (const [name, text] of DOCS) {
+  check(text.includes('same sample findings') === canned, `${name}: demo-mode description disagrees with parser-helper.js (canned output: ${canned})`);
+}
+
+// 3. While the extension can reach Google's API, disclose what Google's free tier may do with submitted text.
+const hosts = [...(manifest.host_permissions || []), ...(manifest.optional_host_permissions || [])];
+if (hosts.some((host) => host.includes('generativelanguage.googleapis.com'))) {
+  for (const [name, text] of [['index.html', html], ['PRIVACY.md', privacy]]) {
+    check(text.includes('improve its products'), `${name} must disclose that Google's free tier may use submitted content to improve its products`);
+  }
+  check(/ai\.google\.dev\/\S*terms/.test(privacy), "PRIVACY.md must link to Google's terms for the API");
+}
+
+// 4. No absolute student-data promises and no compliance claims.
+for (const [name, text] of DOCS) {
+  check(!/\b(?:no|never|zero)\b[^.]*\b(?:PII|personally identifiable)/i.test(text), `${name} makes an absolute no-PII promise`);
+  check(!/FERPA[- ]compli/i.test(text), `${name} claims FERPA compliance`);
+}
+
+// 5. README's crawler bullet may name only Canvas resources the shipped crawler requests.
+const requested = [];
+await crawlCanvasCourse('https://school.instructure.com', '1', async (url) => {
+  requested.push(url);
+  return { ok: true, status: 200, headers: { get: () => null }, json: async () => (/\/courses\/1(\?|$)/.test(url) ? {} : []) };
+});
+const bullet = readme.split('\n').find((l) => l.includes('Zero Developer Tokens Required'));
+check(bullet, "README.md's course-crawler bullet (Zero Developer Tokens Required) is missing");
+for (const [word, endpoint] of [[/\bsyllab/i, 'syllabus_body'], [/\bassignment/i, '/assignments'], [/\bmodule/i, '/modules'], [/\bdiscussion/i, '/discussion_topics'], [/\bquiz/i, '/quizzes']]) {
+  if (bullet && word.test(bullet)) check(requested.some((u) => u.includes(endpoint)), `README.md's crawler bullet claims content the crawler never requests (${endpoint})`);
+}
+
+// 6. Each student-record page the docs say is not read must be refused: neither
+//    its text nor its title may come back, even when every content selector
+//    matches and on a custom Canvas domain, so the guard (not a missing
+//    selector) is what holds.
+const { extractContentFromDOM } = loadContentScript('content/extractor.js');
+const secret = 'Jane Student 92%';
+const el = { innerText: secret, textContent: secret };
+const doc = { title: secret, querySelector: () => el, querySelectorAll: () => [el], body: el };
+const note = (html.match(/<div class="privacy-note">([\s\S]*?)<\/div>/) || [])[1];
+check(note, 'index.html has no privacy note');
+const at = privacy.indexOf('## idstack Chrome Extension');
+check(at >= 0, 'PRIVACY.md has no "## idstack Chrome Extension" section');
+const policy = at >= 0 ? privacy.slice(at) : '';
+for (const [word, url] of [
+  [/Gradebook/, 'https://canvas.example.edu/courses/1/gradebook'],
+  [/SpeedGrader/, 'https://canvas.example.edu/courses/1/gradebook/speed_grader?assignment_id=1'],
+  [/\bgrades\b/, 'https://canvas.example.edu/courses/1/grades/7'],
+  [/People/, 'https://canvas.example.edu/courses/1/users'],
+  [/People/, 'https://canvas.example.edu/courses/1/users/42'],
+  [/\bgroups\b/, 'https://canvas.example.edu/courses/1/groups'],
+  [/\bsubmissions\b/, 'https://canvas.example.edu/courses/1/assignments/5/submissions/7'],
+  [/Inbox/, 'https://canvas.example.edu/conversations'],
+  [/\bdiscussion/i, 'https://canvas.example.edu/courses/1/discussion_topics/1'],
+]) {
+  for (const [name, text] of [['side-panel privacy note', note || ''], ['PRIVACY.md', policy]]) {
+    if (!word.test(text)) continue;
+    const r = extractContentFromDOM(doc, url);
+    check(!JSON.stringify([r.title, r.content]).includes(secret), `${name} says ${word.source} pages are not read, but the extractor returned page text or title for ${url}`);
+  }
+}
+
+// 7. Every host the manifest lets the extension reach must be named in the policy, and a
+//    settings field for a Consensus key means the panel's privacy note must mention Consensus.
+//    #89 added api.consensus.app to host_permissions and a key field with neither.
+for (const pattern of manifest.host_permissions || []) {
+  const host = pattern.replace(/^[^:]+:\/\//, '').replace(/^\*\./, '').replace(/\/.*$/, '');
+  check(policy.includes(host), `PRIVACY.md's extension section does not name ${host}, which manifest.json host_permissions lets the extension reach`);
+}
+if (html.includes('id="consensus-api-key-input"')) {
+  check(/Consensus/.test(note || ''), 'side-panel privacy note does not say what a saved Consensus key sends to Consensus');
+}
+
+assert.ok(problems.size === 0, `Extension disclosures disagree with the shipped code:\n  ${[...problems].join('\n  ')}`);
+console.log('✅ Extension privacy and capability disclosures match the shipped code.');
diff --git a/test/test-dossier-compiler.js b/test/test-dossier-compiler.mjs
similarity index 77%
rename from test/test-dossier-compiler.js
rename to test/test-dossier-compiler.mjs
index 8db95e7..7a3b38c 100644
--- a/test/test-dossier-compiler.js
+++ b/test/test-dossier-compiler.mjs
@@ -1,7 +1,5 @@
-const fs = require('fs');
-const path = require('path');
-const assert = require('assert');
-const { compileDossierToMarkdown, compileSingleAuditToMarkdown } = require('../extension/shared/dossier-compiler.cjs');
+import assert from 'node:assert';
+import { compileDossierToMarkdown, compileSingleAuditToMarkdown } from '../extension/shared/dossier-compiler.js';
 
 // Test 1: Single Audit Markdown Compilation
 const singleItem = {
@@ -86,13 +84,4 @@ assert.ok(emptyDossierMd.includes('*No audit materials in dossier.*'));
 const nullDossierMd = compileDossierToMarkdown(null);
 assert.ok(nullDossierMd.includes('*No audit materials in dossier.*'));
 
-// Test 4: Verify storage.js methods exist
-const storagePath = path.join(__dirname, '../extension/shared/storage.js');
-assert.ok(fs.existsSync(storagePath), 'storage.js must exist');
-const storageContent = fs.readFileSync(storagePath, 'utf8');
-assert.ok(storageContent.includes('export async function getDossier()'), 'getDossier export required');
-assert.ok(storageContent.includes('export async function addToDossier('), 'addToDossier export required');
-assert.ok(storageContent.includes('export async function removeFromDossier('), 'removeFromDossier export required');
-assert.ok(storageContent.includes('export async function clearDossier()'), 'clearDossier export required');
-
 console.log('✅ Task 1 Dossier compiler tests passed.');
diff --git a/test/test-evidence-labels.mjs b/test/test-evidence-labels.mjs
new file mode 100644
index 0000000..db7ae6b
--- /dev/null
+++ b/test/test-evidence-labels.mjs
@@ -0,0 +1,98 @@
+// Evidence labels shipped in the extension must agree with evidence/references.md
+// (codes, studies, tiers) and CLAUDE.md (tier meanings). The extension shipped a
+// tier scale that put randomized trials in T2, demo findings citing Biggs (T5,
+// [Alignment-1]) as "[Alignment-3] ... [T2]" and a nonexistent [Cognitive-2],
+// uncoded prompt citations to studies references.md does not contain, and a page
+// prompt asking for a "suggestion" severity. Expected values are parsed from the
+// canonical files, never restated here.
+import assert from 'node:assert';
+import fs from 'node:fs';
+import { TIER_METADATA } from '../extension/shared/evidence-base.js';
+import { getDemoAuditResult, getDemoCourseAuditResult } from '../extension/background/parser-helper.js';
+import { buildAuditPrompt, buildCourseAuditPrompt } from '../extension/shared/prompts.js';
+
+const read = (rel) => fs.readFileSync(new URL(`../${rel}`, import.meta.url), 'utf8');
+const SEVERITIES = ['critical', 'warning', 'info'];
+const SOURCES = ['extension/shared/prompts.js', 'extension/background/parser-helper.js', 'extension/shared/evidence-base.js'];
+const CODE = /\[([A-Z][A-Za-z]+-\d+)\](?: \[(T[1-5])\])?/g; // capitalized, so arr[i-1] is no code
+// "[Code-N] Author (Year) [Tn]", "Author et al. Year [Tn", "Author (Year)": code and tier optional.
+const STUDY = /(?:\[([A-Z][A-Za-z]+-\d+)\] )?([A-Z][A-Za-z'-]+)(?: et al\.)? \(?((?:19|20)\d{2})\)?(?: \[(T[1-5])\b)?/g;
+
+const refsText = read('evidence/references.md');
+const refs = {}; // 'Alignment-1' -> { line, tier }
+for (const m of refsText.matchAll(/^- \[([A-Za-z]+-\d+)\] (.+) (T[1-5])\s*$/gm)) refs[m[1]] = { line: m[2], tier: m[3] };
+const tierDefs = {}; // references.md "Definition" column
+for (const m of refsText.matchAll(/^\| (T[1-5]) \| ([^|]+?) \|/gm)) tierDefs[m[1]] = m[2];
+const tierMeanings = {}; // CLAUDE.md "Evidence standards" "Meaning" column
+for (const m of read('CLAUDE.md').matchAll(/^\| (T[1-5]) \| ([^|]+?) \|\s*$/gm)) tierMeanings[m[1]] = m[2];
+// A parse that finds nothing would make every check below vacuous.
+assert.ok(Object.keys(refs).length >= 100, 'parsed too few references from evidence/references.md');
+assert.strictEqual(Object.keys(tierDefs).length, 5, 'references.md tier table not parsed');
+assert.strictEqual(Object.keys(tierMeanings).length, 5, 'CLAUDE.md tier table not parsed');
+
+// Collect every disagreement so one run names them all.
+const problems = new Set();
+const check = (ok, msg) => { if (!ok) problems.add(msg); };
+const names = (line, author, year) => line.includes(author) && line.includes(`(${year})`);
+
+// A code must be a references.md entry at the stated tier, naming the stated study.
+function cite(where, code, tier, author, year) {
+  const ref = refs[code];
+  if (!ref) return check(false, `${where}: [${code}] is not in evidence/references.md`);
+  if (tier) check(tier === ref.tier, `${where}: [${code}] is ${ref.tier} in references.md, not ${tier}`);
+  if (author) check(names(ref.line, author, year), `${where}: ${author} (${year}) is not [${code}] ${ref.line}`);
+}
+// An uncoded "Author (Year)" must still be a references.md entry, at the stated tier.
+function study(where, author, year, tier) {
+  const entries = Object.values(refs).filter((r) => names(r.line, author, year));
+  check(entries.length, `${where}: ${author} (${year}) is not in evidence/references.md`);
+  if (entries.length && tier) check(entries.some((r) => r.tier === tier), `${where}: ${author} (${year}) is ${entries[0].tier} in references.md, not ${tier}`);
+}
+
+// 1. Every [Code-N] and every "Author (Year)" in the shipped files is a
+//    references.md entry, at whatever tier the text states beside it.
+for (const rel of SOURCES) {
+  const text = read(rel);
+  for (const m of text.matchAll(CODE)) cite(rel, m[1], m[2]);
+  for (const m of text.matchAll(STUDY)) {
+    if (m[1]) cite(rel, m[1], m[4], m[2], m[3]);
+    else study(rel, m[2], m[3], m[4]);
+  }
+}
+
+// 2. The tier scale restates the canonical tables.
+for (const t of Object.keys(tierDefs)) {
+  check(TIER_METADATA[t]?.description === tierDefs[t], `TIER_METADATA.${t}.description must be references.md's "${tierDefs[t]}"`);
+  check(TIER_METADATA[t]?.label === tierMeanings[t], `TIER_METADATA.${t}.label must be CLAUDE.md's "${tierMeanings[t]}"`);
+}
+
+// 3. Demo findings: canonical severity; the citation's code is at the finding's
+//    tier; the evidence names that code's study and states the same tier.
+for (const [where, result] of [['page demo', getDemoAuditResult({})], ['course demo', getDemoCourseAuditResult({})]]) {
+  for (const f of result.findings) {
+    const at = `${where} "${f.citation}"`;
+    check(SEVERITIES.includes(f.severity), `${at}: severity "${f.severity}" is not ${SEVERITIES.join('|')}`);
+    const code = (f.citation.match(/^\[([A-Za-z]+-\d+)\]/) || [])[1];
+    const named = f.evidence.match(/^([A-Z][A-Za-z'-]+)[^(]*\((\d{4})\)/) || [];
+    check(named[1], `${at}: evidence names no "Author (Year)"`);
+    cite(at, code, f.tier, named[1], named[2]);
+    for (const t of f.evidence.matchAll(/\[(T[1-5])\b/g)) check(t[1] === f.tier, `${at}: evidence says ${t[1]}, finding says ${f.tier}`);
+  }
+}
+
+// 4. Prompts: one severity vocabulary; every tiered citation carries its code;
+//    both define the canonical tier scale.
+for (const [where, prompt] of [['page prompt', buildAuditPrompt({ title: 't' })], ['course prompt', buildCourseAuditPrompt({})]]) {
+  const vocab = ((prompt.match(/"severity": (.*)/) || [])[1] || '').match(/[a-z]+/g) || [];
+  check(vocab.join('|') === SEVERITIES.join('|'), `${where}: severity vocabulary is ${vocab.join('|')}, not ${SEVERITIES.join('|')}`);
+  for (const m of prompt.matchAll(STUDY)) check(m[1] || !m[4], `${where}: "${m[0]}" states a tier without a references.md code`);
+  for (const t of Object.keys(tierDefs)) check(prompt.includes(`${t}: ${tierDefs[t]}`), `${where} must define ${t} as "${tierDefs[t]}"`);
+}
+
+// 5. The side panel styles exactly the canonical severities (rules and tokens).
+const css = read('extension/sidepanel/sidepanel.css');
+const styled = [...new Set([...css.matchAll(/(?:\.severity|--sev(?:erity)?)-([a-z]+)/g)].map((m) => m[1]))].sort();
+check(styled.join('|') === [...SEVERITIES].sort().join('|'), `sidepanel.css styles severities ${styled.join('|')}, not ${SEVERITIES.join('|')}`);
+
+assert.ok(problems.size === 0, `Extension evidence labels disagree with evidence/references.md or CLAUDE.md:\n  ${[...problems].join('\n  ')}`);
+console.log('✅ Evidence label tests passed.');
diff --git a/test/test-extension.sh b/test/test-extension.sh
index a5e5f49..571d5d3 100755
--- a/test/test-extension.sh
+++ b/test/test-extension.sh
@@ -3,14 +3,26 @@ set -euo pipefail
 
 DIR="$(cd "$(dirname "$0")" && pwd)"
 
+# The tests import the shipped extension modules directly. extension/ has no
+# package.json (Chrome needs none), so Node must detect ES module syntax in .js
+# files itself: on by default from Node 20.19 and 22.7.
+if ! node -e 'const [a,b]=process.versions.node.split(".").map(Number);process.exit(a>22||(a===22&&b>=7)||(a===20&&b>=19)?0:1)'; then
+  echo "test-extension.sh needs Node >= 20.19 or >= 22.7 (found $(node --version))" >&2
+  exit 1
+fi
+
+# Test the shipped file, never a copy: hand-maintained .cjs twins let the tests
+# pass against code the extension never ran.
+if find "$DIR/../extension" -name '*.cjs' | grep .; then
+  echo "extension/ must not contain .cjs copies; tests import the shipped .js" >&2
+  exit 1
+fi
+
 echo "==> Running idstack Chrome Extension test suite..."
 node "$DIR/test-manifest.js"
-node "$DIR/test-prompts.js"
-node "$DIR/test-extractor.js"
-node "$DIR/test-crawler.js"
-node "$DIR/test-service-worker.js"
-node "$DIR/test-dossier-compiler.js"
 node "$DIR/test-sidepanel-dom.js"
-node "$DIR/test-sidepanel-logic.js"
+for t in "$DIR"/test-*.mjs; do
+  node "$t"
+done
 
 echo "==> All Chrome Extension tests passed!"
diff --git a/test/test-extractor.js b/test/test-extractor.js
deleted file mode 100644
index 9081df9..0000000
--- a/test/test-extractor.js
+++ /dev/null
@@ -1,338 +0,0 @@
-const assert = require('assert');
-
-// Support JSDOM if available, or fallback to built-in DOM mock for standalone node environments
-let JSDOM;
-try {
-  JSDOM = require('jsdom').JSDOM;
-} catch {
-  class MockElement {
-    constructor(tagName = 'div', attrs = {}, text = '') {
-      this.tagName = tagName.toLowerCase();
-      this.attrs = attrs;
-      this.children = [];
-      this._text = text;
-    }
-    get innerText() {
-      if (this.children.length > 0) {
-        return this.children.map(c => typeof c === 'string' ? c : c.innerText).join(' ').trim();
-      }
-      return this._text.trim();
-    }
-    get textContent() {
-      return this.innerText;
-    }
-    querySelector(selector) {
-      const selectors = selector.split(',').map(s => s.trim());
-      for (const sel of selectors) {
-        const found = this._querySingle(sel);
-        if (found) return found;
-      }
-      return null;
-    }
-    _querySingle(selector) {
-      const parts = selector.split(/\s+/);
-      let current = [this];
-      for (const part of parts) {
-        const next = [];
-        for (const el of current) {
-          next.push(...el._findDescendants(part));
-        }
-        current = next;
-        if (current.length === 0) return null;
-      }
-      return current[0] || null;
-    }
-    _findDescendants(part) {
-      const results = [];
-      const check = (el) => {
-        if (el !== this) {
-          let match = true;
-          if (part.startsWith('#')) {
-            if (el.attrs.id !== part.slice(1)) match = false;
-          } else if (part.startsWith('.')) {
-            const requiredClasses = part.split('.').filter(Boolean);
-            const elClasses = (el.attrs.class || '').split(/\s+/).filter(Boolean);
-            if (!requiredClasses.every(cls => elClasses.includes(cls))) match = false;
-          } else if (part.startsWith('[')) {
-            const m = part.match(/\[([a-zA-Z0-9_-]+)(?:="([^"]*)")?\]/);
-            if (m) {
-              const [, attrName, attrVal] = m;
-              if (attrVal !== undefined) {
-                if (el.attrs[attrName] !== attrVal) match = false;
-              } else {
-                if (!(attrName in el.attrs)) match = false;
-              }
-            }
-          } else {
-            if (el.tagName !== part.toLowerCase()) match = false;
-          }
-          if (match) results.push(el);
-        }
-        for (const child of el.children) {
-          if (typeof child !== 'string') check(child);
-        }
-      };
-      check(this);
-      return results;
-    }
-  }
-
-  function parseHtmlToTree(html) {
-    const root = new MockElement('root');
-    const stack = [root];
-    const tagRegex = /<(\/)?([a-zA-Z0-9]+)([^>]*)>|([^<]+)/g;
-    let match;
-
-    while ((match = tagRegex.exec(html)) !== null) {
-      const [full, isClose, tagName, attrStr, text] = match;
-      if (text) {
-        const trimmed = text.trim();
-        if (trimmed) {
-          const current = stack[stack.length - 1];
-          current.children.push(trimmed);
-        }
-      } else if (tagName) {
-        if (isClose) {
-          if (stack.length > 1 && stack[stack.length - 1].tagName === tagName.toLowerCase()) {
-            stack.pop();
-          }
-        } else {
-          const attrs = {};
-          if (attrStr) {
-            const attrRegex = /([a-zA-Z0-9_-]+)(?:=["']([^"']*)["'])?/g;
-            let attrMatch;
-            while ((attrMatch = attrRegex.exec(attrStr)) !== null) {
-              attrs[attrMatch[1]] = attrMatch[2] || '';
-            }
-          }
-          const el = new MockElement(tagName, attrs);
-          stack[stack.length - 1].children.push(el);
-          if (!['meta', 'link', 'img', 'br', 'hr', 'input'].includes(tagName.toLowerCase())) {
-            stack.push(el);
-          }
-        }
-      }
-    }
-    return root;
-  }
-
-  class SimpleJSDOM {
-    constructor(html, { url = 'http://localhost' } = {}) {
-      const tree = parseHtmlToTree(html);
-      const titleEl = tree.querySelector('title');
-      const titleText = titleEl ? titleEl.innerText : '';
-      const bodyEl = tree.querySelector('body') || tree;
-
-      this.window = {
-        location: { href: url },
-        document: {
-          title: titleText,
-          body: bodyEl,
-          querySelector: (sel) => tree.querySelector(sel)
-        }
-      };
-    }
-  }
-  JSDOM = SimpleJSDOM;
-}
-
-const { detectPageType, extractContentFromDOM, extractPageContent } = require('../extension/content/extractor-core.cjs');
-
-// Test 1: Canvas Assignment fixture
-const canvasAssignmentHTML = `
-  <html>
-    <head><title>Module 3: Enzymes Assignment</title></head>
-    <body>
-      <div id="assignment_show">
-        <h1 class="title">Enzymes Lab Analysis</h1>
-        <div class="description user_content">
-          <p>Read chapter 4 and answer the 5 review questions.</p>
-        </div>
-      </div>
-    </body>
-  </html>
-`;
-const domAssignment = new JSDOM(canvasAssignmentHTML, { url: 'https://canvas.instructure.com/courses/101/assignments/202' });
-const extractedAssignment = extractContentFromDOM(domAssignment.window.document, domAssignment.window.location.href);
-
-assert.strictEqual(extractedAssignment.pageType, 'Canvas Assignment');
-assert.strictEqual(extractedAssignment.title, 'Enzymes Lab Analysis');
-assert.ok(extractedAssignment.content.includes('Read chapter 4'));
-assert.strictEqual(extractedAssignment.wordCount, 9);
-
-// Test 2: Canvas Syllabus fixture
-const canvasSyllabusHTML = `
-  <html>
-    <head><title>Course Syllabus - CS 101</title></head>
-    <body>
-      <h1 class="page-title">CS 101 Syllabus</h1>
-      <div id="syllabusContainer">
-        <p>This course covers algorithms and data structures.</p>
-      </div>
-    </body>
-  </html>
-`;
-const domSyllabus = new JSDOM(canvasSyllabusHTML, { url: 'https://canvas.instructure.com/courses/101/assignments/syllabus' });
-const extractedSyllabus = extractContentFromDOM(domSyllabus.window.document, domSyllabus.window.location.href);
-
-assert.strictEqual(extractedSyllabus.pageType, 'Canvas Syllabus');
-assert.strictEqual(extractedSyllabus.title, 'CS 101 Syllabus');
-assert.ok(extractedSyllabus.content.includes('algorithms and data structures'));
-
-// Test 3: Canvas Modules fixture
-const canvasModulesHTML = `
-  <html>
-    <head><title>Course Modules - BIO 200</title></head>
-    <body>
-      <h1 class="page-title">BIO 200 Modules</h1>
-      <div id="modules">
-        <span class="module-item-title">Week 1: Cellular Respiration</span>
-      </div>
-    </body>
-  </html>
-`;
-const domModules = new JSDOM(canvasModulesHTML, { url: 'https://canvas.instructure.com/courses/101/modules' });
-const extractedModules = extractContentFromDOM(domModules.window.document, domModules.window.location.href);
-
-assert.strictEqual(extractedModules.pageType, 'Canvas Modules');
-assert.strictEqual(extractedModules.title, 'BIO 200 Modules');
-assert.ok(extractedModules.content.includes('Week 1: Cellular Respiration'));
-
-// Test 4: Canvas Rubrics fixture
-const canvasRubricsHTML = `
-  <html>
-    <head><title>Course Rubrics</title></head>
-    <body>
-      <h1 class="page-title">Grading Criteria</h1>
-      <div id="rubrics">
-        <p>Criteria 1: Clarity of argument (10 pts)</p>
-      </div>
-    </body>
-  </html>
-`;
-const domRubrics = new JSDOM(canvasRubricsHTML, { url: 'https://canvas.instructure.com/courses/101/rubrics' });
-const extractedRubrics = extractContentFromDOM(domRubrics.window.document, domRubrics.window.location.href);
-
-assert.strictEqual(extractedRubrics.pageType, 'Canvas Rubric');
-assert.strictEqual(extractedRubrics.title, 'Grading Criteria');
-
-// Test 5: Canvas Generic LMS Page
-const canvasGenericHTML = `
-  <html>
-    <head><title>Course Overview</title></head>
-    <body>
-      <h1 class="page-title">Welcome to PHY 101</h1>
-      <div class="show-content user_content">
-        <p>General introduction to physics principles.</p>
-      </div>
-    </body>
-  </html>
-`;
-const domGeneric = new JSDOM(canvasGenericHTML, { url: 'https://canvas.instructure.com/courses/101/pages/overview' });
-const extractedGeneric = extractContentFromDOM(domGeneric.window.document, domGeneric.window.location.href);
-
-assert.strictEqual(extractedGeneric.pageType, 'Canvas LMS Page');
-assert.strictEqual(extractedGeneric.title, 'Welcome to PHY 101');
-assert.ok(extractedGeneric.content.includes('physics principles'));
-
-// Test 6: Google Docs fixture
-const gdocHTML = `
-  <html>
-    <head><title>Instructional Design Principles - Google Docs</title></head>
-    <body>
-      <div class="kix-appview-editor">
-        <p>Course Outline: 1. Assessment Design 2. Cognitive Load Management</p>
-      </div>
-    </body>
-  </html>
-`;
-const domGdoc = new JSDOM(gdocHTML, { url: 'https://docs.google.com/document/d/1BxiMVs0XRA5nFMdKvBdBZjgmUUqptlbs74OgvE2upms/edit' });
-const extractedGdoc = extractContentFromDOM(domGdoc.window.document, domGdoc.window.location.href);
-
-assert.strictEqual(extractedGdoc.pageType, 'Google Doc Syllabus');
-assert.ok(extractedGdoc.content.includes('Assessment Design'));
-
-// Test 7: Generic Web Page fixture
-const webHTML = `
-  <html>
-    <head><title>Open Courseware Syllabus</title></head>
-    <body>
-      <main>
-        <h1>Introduction to Machine Learning</h1>
-        <p>Course schedule and grading policy.</p>
-      </main>
-    </body>
-  </html>
-`;
-const domWeb = new JSDOM(webHTML, { url: 'https://ocw.mit.edu/syllabus/intro-ml' });
-const extractedWeb = extractContentFromDOM(domWeb.window.document, domWeb.window.location.href);
-
-assert.strictEqual(extractedWeb.pageType, 'Web Syllabus / Course Page');
-assert.strictEqual(extractedWeb.title, 'Open Courseware Syllabus');
-assert.ok(extractedWeb.content.includes('Machine Learning'));
-
-// Test 8: Content length cap at 15000 characters
-const longText = 'word '.repeat(4000); // 20,000 characters
-const longHTML = `<html><head><title>Long Document</title></head><body><main><p>${longText}</p></main></body></html>`;
-const domLong = new JSDOM(longHTML, { url: 'https://example.com/long' });
-const extractedLong = extractContentFromDOM(domLong.window.document, domLong.window.location.href);
-
-assert.strictEqual(extractedLong.content.length, 15000, 'Content should be capped at 15000 chars');
-
-// Test 9: detectPageType helper directly
-assert.strictEqual(detectPageType('https://canvas.instructure.com/courses/1/assignments/2', { querySelector: (s) => s === '#assignment_show' ? {} : null }), 'Canvas Assignment');
-assert.strictEqual(detectPageType('https://docs.google.com/document/d/123', {}), 'Google Doc Syllabus');
-assert.strictEqual(detectPageType('https://harvard.edu/syllabus', {}), 'Web Syllabus / Course Page');
-
-// Test 10: Chrome Runtime onMessage listener handling
-let messageListener = null;
-global.chrome = {
-  runtime: {
-    onMessage: {
-      addListener: (fn) => {
-        messageListener = fn;
-      }
-    }
-  }
-};
-global.window = domAssignment.window;
-global.document = domAssignment.window.document;
-
-delete require.cache[require.resolve('../extension/content/extractor.js')];
-require('../extension/content/extractor.js');
-
-assert.ok(typeof messageListener === 'function', 'chrome.runtime.onMessage listener should be registered');
-let responseData = null;
-messageListener({ action: 'EXTRACT_CONTENT' }, {}, (data) => {
-  responseData = data;
-});
-assert.ok(responseData, 'Response data must be returned on EXTRACT_CONTENT action');
-assert.strictEqual(responseData.pageType, 'Canvas Assignment');
-assert.strictEqual(responseData.title, 'Enzymes Lab Analysis');
-
-// Clean up globals
-delete global.chrome;
-delete global.window;
-delete global.document;
-
-// Test 11: Canvas Course Root Detection
-const { detectCourseContext } = require('../extension/content/extractor-core.cjs');
-const rootCtx = detectCourseContext('https://canvas.instructure.com/courses/987654', 'Biology 101');
-assert.strictEqual(rootCtx.isCourseRoot, true);
-assert.strictEqual(rootCtx.courseId, '987654');
-assert.strictEqual(rootCtx.origin, 'https://canvas.instructure.com');
-
-const modulesCtx = detectCourseContext('https://canvas.instructure.com/courses/987654/modules', 'Modules');
-assert.strictEqual(modulesCtx.isCourseRoot, true);
-assert.strictEqual(modulesCtx.courseId, '987654');
-assert.strictEqual(modulesCtx.origin, 'https://canvas.instructure.com');
-
-const queryParamCtx = detectCourseContext('https://canvas.instructure.com/courses/987654/modules?view=feed', 'Modules Feed');
-assert.strictEqual(queryParamCtx.isCourseRoot, true);
-assert.strictEqual(queryParamCtx.courseId, '987654');
-
-const subpageCtx = detectCourseContext('https://canvas.instructure.com/courses/987654/assignments/123', 'Lab 1');
-assert.strictEqual(subpageCtx.isCourseRoot, false);
-assert.strictEqual(subpageCtx.courseId, '987654');
-
-console.log('✅ Task 3 extractor tests passed.');
diff --git a/test/test-extractor.mjs b/test/test-extractor.mjs
new file mode 100644
index 0000000..a3ecedd
--- /dev/null
+++ b/test/test-extractor.mjs
@@ -0,0 +1,277 @@
+import assert from 'node:assert';
+import { readFileSync } from 'node:fs';
+import vm from 'node:vm';
+import { createDocument, installChrome, installFetch, loadContentScript, resetGlobals } from './extension-harness.mjs';
+
+// The shipped content script: a classic script, loaded exactly as Chrome gets it.
+const { detectPageType, extractContentFromDOM, extractPageContent } = loadContentScript('content/extractor.js');
+const page = (html, url) => ({ url, document: createDocument(html) });
+
+// Test 1: Canvas Assignment fixture
+const canvasAssignmentHTML = `
+  <html>
+    <head><title>Module 3: Enzymes Assignment</title></head>
+    <body>
+      <div id="assignment_show">
+        <h1 class="title">Enzymes Lab Analysis</h1>
+        <div class="description user_content">
+          <p>Read chapter 4 and answer the 5 review questions.</p>
+        </div>
+      </div>
+    </body>
+  </html>
+`;
+const domAssignment = page(canvasAssignmentHTML, 'https://canvas.instructure.com/courses/101/assignments/202');
+const extractedAssignment = extractContentFromDOM(domAssignment.document, domAssignment.url);
+
+assert.strictEqual(extractedAssignment.pageType, 'Canvas Assignment');
+assert.strictEqual(extractedAssignment.title, 'Enzymes Lab Analysis');
+assert.ok(extractedAssignment.content.includes('Read chapter 4'));
+assert.strictEqual(extractedAssignment.wordCount, 9);
+
+// Test 2: Canvas Syllabus fixture
+const canvasSyllabusHTML = `
+  <html>
+    <head><title>Course Syllabus - CS 101</title></head>
+    <body>
+      <h1 class="page-title">CS 101 Syllabus</h1>
+      <div id="syllabusContainer">
+        <p>This course covers algorithms and data structures.</p>
+      </div>
+    </body>
+  </html>
+`;
+const domSyllabus = page(canvasSyllabusHTML, 'https://canvas.instructure.com/courses/101/assignments/syllabus');
+const extractedSyllabus = extractContentFromDOM(domSyllabus.document, domSyllabus.url);
+
+assert.strictEqual(extractedSyllabus.pageType, 'Canvas Syllabus');
+assert.strictEqual(extractedSyllabus.title, 'CS 101 Syllabus');
+assert.ok(extractedSyllabus.content.includes('algorithms and data structures'));
+
+// Test 3: Canvas Modules fixture
+const canvasModulesHTML = `
+  <html>
+    <head><title>Course Modules - BIO 200</title></head>
+    <body>
+      <h1 class="page-title">BIO 200 Modules</h1>
+      <div id="modules">
+        <span class="module-item-title">Week 1: Cellular Respiration</span>
+        <span class="module-item-title">Lab 1: Measuring Oxygen Uptake</span>
+      </div>
+    </body>
+  </html>
+`;
+const domModules = page(canvasModulesHTML, 'https://canvas.instructure.com/courses/101/modules');
+const extractedModules = extractContentFromDOM(domModules.document, domModules.url);
+
+assert.strictEqual(extractedModules.pageType, 'Canvas Modules');
+assert.strictEqual(extractedModules.title, 'BIO 200 Modules');
+assert.ok(extractedModules.content.includes('Week 1: Cellular Respiration'));
+assert.ok(extractedModules.content.includes('Lab 1: Measuring Oxygen Uptake'), 'Modules must read every item title, not just the first');
+
+// A course home in Modules view: no /modules in the URL, and Canvas's container is #context_modules.
+const domCourseHome = page(`<html><head><title>BIO 200</title></head><body><div id="context_modules"><span class="module-item-title">Week 1: Cellular Respiration</span><span class="module-item-title">Lab 1: Measuring Oxygen Uptake</span></div></body></html>`, 'https://canvas.instructure.com/courses/101');
+const extractedCourseHome = extractContentFromDOM(domCourseHome.document, domCourseHome.url);
+assert.strictEqual(extractedCourseHome.pageType, 'Canvas Modules', 'course home in Modules view is a Modules page');
+assert.ok(extractedCourseHome.content.includes('Lab 1: Measuring Oxygen Uptake'), 'course home Modules view reads every item title');
+
+// Test 4: Canvas Rubrics fixture
+const canvasRubricsHTML = `
+  <html>
+    <head><title>Course Rubrics</title></head>
+    <body>
+      <h1 class="page-title">Grading Criteria</h1>
+      <div id="rubrics">
+        <p>Criteria 1: Clarity of argument (10 pts)</p>
+      </div>
+    </body>
+  </html>
+`;
+const domRubrics = page(canvasRubricsHTML, 'https://canvas.instructure.com/courses/101/rubrics');
+const extractedRubrics = extractContentFromDOM(domRubrics.document, domRubrics.url);
+
+assert.strictEqual(extractedRubrics.pageType, 'Canvas Rubric');
+assert.strictEqual(extractedRubrics.title, 'Grading Criteria');
+assert.ok(extractedRubrics.content.includes('Criteria 1'), 'rubric text must be read without a document.body fallback');
+
+// Test 5: Canvas Generic LMS Page
+const canvasGenericHTML = `
+  <html>
+    <head><title>Course Overview</title></head>
+    <body>
+      <h1 class="page-title">Welcome to PHY 101</h1>
+      <div class="show-content user_content">
+        <p>General introduction to physics principles.</p>
+      </div>
+    </body>
+  </html>
+`;
+const domGeneric = page(canvasGenericHTML, 'https://canvas.instructure.com/courses/101/pages/overview');
+const extractedGeneric = extractContentFromDOM(domGeneric.document, domGeneric.url);
+
+assert.strictEqual(extractedGeneric.pageType, 'Canvas LMS Page');
+assert.strictEqual(extractedGeneric.title, 'Welcome to PHY 101');
+assert.ok(extractedGeneric.content.includes('physics principles'));
+
+// Test 6: Google Docs fixture
+const gdocHTML = `
+  <html>
+    <head><title>Instructional Design Principles - Google Docs</title></head>
+    <body>
+      <div class="kix-appview-editor">
+        <p>Course Outline: 1. Assessment Design 2. Cognitive Load Management</p>
+      </div>
+    </body>
+  </html>
+`;
+const domGdoc = page(gdocHTML, 'https://docs.google.com/document/d/1BxiMVs0XRA5nFMdKvBdBZjgmUUqptlbs74OgvE2upms/edit');
+const extractedGdoc = extractContentFromDOM(domGdoc.document, domGdoc.url);
+
+assert.strictEqual(extractedGdoc.pageType, 'Google Doc Syllabus');
+assert.ok(extractedGdoc.content.includes('Assessment Design'));
+
+// Test 7: Generic Web Page fixture
+const webHTML = `
+  <html>
+    <head><title>Open Courseware Syllabus</title></head>
+    <body>
+      <main>
+        <h1>Introduction to Machine Learning</h1>
+        <p>Course schedule and grading policy.</p>
+      </main>
+    </body>
+  </html>
+`;
+const domWeb = page(webHTML, 'https://ocw.mit.edu/syllabus/intro-ml');
+const extractedWeb = extractContentFromDOM(domWeb.document, domWeb.url);
+
+assert.strictEqual(extractedWeb.pageType, 'Web Syllabus / Course Page');
+assert.strictEqual(extractedWeb.title, 'Open Courseware Syllabus');
+assert.ok(extractedWeb.content.includes('Machine Learning'));
+
+// Test 8: Content length cap at 15000 characters
+const longText = 'word '.repeat(4000); // 20,000 characters
+const longHTML = `<html><head><title>Long Document</title></head><body><main><p>${longText}</p></main></body></html>`;
+const domLong = page(longHTML, 'https://example.com/long');
+const extractedLong = extractContentFromDOM(domLong.document, domLong.url);
+
+assert.strictEqual(extractedLong.content.length, 15000, 'Content should be capped at 15000 chars');
+
+// Test 9: detectPageType helper directly
+assert.strictEqual(detectPageType('https://canvas.instructure.com/courses/1/assignments/2', { querySelector: (s) => s === '#assignment_show' ? {} : null }), 'Canvas Assignment');
+assert.strictEqual(detectPageType('https://docs.google.com/document/d/123', {}), 'Google Doc Syllabus');
+assert.strictEqual(detectPageType('https://harvard.edu/syllabus', {}), 'Web Syllabus / Course Page');
+
+// Test 10: executeScript({ files }) contract. The side panel injects the shipped
+// extractor.js and reads the script's completion value as the InjectionResult.
+// It re-injects on every refresh into the same isolated world, so a second run
+// must work too (a top-level let/const/class would throw on redeclaration).
+const extractorUrl = new URL('../extension/content/extractor.js', import.meta.url);
+const extractorSrc = readFileSync(extractorUrl, 'utf8');
+const h = installChrome();
+const isolatedWorld = vm.createContext({
+  URL,
+  chrome: h.chrome,
+  document: domAssignment.document,
+  window: { location: { href: domAssignment.url } }
+});
+for (const run of ['first', 'second']) {
+  // extractPageContent is async; Chrome awaits a Promise completion value.
+  const injected = await vm.runInContext(extractorSrc, isolatedWorld, { filename: extractorUrl.pathname });
+  assert.ok(injected && injected.pageType === 'Canvas Assignment', `${run} injection must evaluate to the extraction payload`);
+  assert.strictEqual(injected.title, 'Enzymes Lab Analysis');
+}
+assert.strictEqual(h.listeners.message.length, 0, 'extractor must not register a persistent runtime.onMessage listener');
+resetGlobals();
+
+// Test 12: student-record pages are refused, and Canvas never falls back to the whole page (F11)
+const studentRecordHTML = `
+  <html>
+    <head><title>Jane Doe: BIO 200</title></head>
+    <body>
+      <div id="content" role="main">
+        <h1>Jane Doe</h1>
+        <p>Jane Doe jane.doe@example.edu Section 01 Final grade 72%</p>
+      </div>
+    </body>
+  </html>
+`;
+for (const url of [
+  'https://canvas.instructure.com/courses/101/gradebook',
+  'https://canvas.instructure.com/courses/101/gradebook/speed_grader?assignment_id=5',
+  'https://canvas.instructure.com/courses/101/grades/7',
+  'https://canvas.instructure.com/courses/101/users/7',
+  'https://canvas.instructure.com/courses/101/groups',
+  'https://canvas.instructure.com/courses/101/discussion_topics/9',
+  'https://canvas.instructure.com/courses/101/assignments/5/submissions/7',
+  'https://canvas.instructure.com/conversations',
+  // Canvas on a custom domain outside /courses/N is not detected as Canvas; the guard still applies.
+  'https://canvas.example.edu/conversations',
+  'https://canvas.example.edu/groups/12',
+  'https://canvas.example.edu/users/7',
+  'https://canvas.example.edu/accounts/1/users'
+]) {
+  const r = extractContentFromDOM(createDocument(studentRecordHTML), url);
+  assert.strictEqual(r.content, '', `${url}: student-record page must not be read`);
+  assert.ok(r.emptyReason, `${url}: must say why nothing was read`);
+  assert.ok(!JSON.stringify(r).includes('Jane Doe'), `${url}: student name leaked into the payload`);
+}
+
+// An unlisted Canvas view gets '' rather than the whole page.
+const quizHistoryUrl = 'https://canvas.instructure.com/courses/101/quizzes/5/history?quiz_submission_id=3';
+assert.strictEqual(extractContentFromDOM(createDocument(studentRecordHTML), quizHistoryUrl).content, '', 'Canvas must not fall back to document.body');
+
+// Custom-domain Canvas course pages are Canvas (no [role=main] or body read).
+const customCourseUrl = 'https://canvas.example.edu/courses/101/quizzes/5/history?quiz_submission_id=3';
+const extractedCustomCourse = extractContentFromDOM(createDocument(studentRecordHTML), customCourseUrl);
+assert.ok(extractedCustomCourse.pageType.startsWith('Canvas'), 'custom-domain /courses/N URLs are Canvas');
+assert.strictEqual(extractedCustomCourse.content, '', 'custom-domain Canvas must not fall back to the whole page');
+
+// Non-Canvas /courses/<slug> sites keep generic extraction.
+const ocwUrl = 'https://ocw.mit.edu/courses/6-006-introduction-to-algorithms-spring-2020/pages/syllabus/';
+const extractedOcw = extractContentFromDOM(createDocument(webHTML), ocwUrl);
+assert.strictEqual(extractedOcw.pageType, 'Web Syllabus / Course Page', 'non-numeric /courses/ URLs are not Canvas');
+assert.ok(extractedOcw.content.includes('Machine Learning'));
+
+// No URL at all (url defaults to '' outside a page): the guard must not throw
+// (new URL('') does) and must not refuse the page.
+let extractedNoUrl;
+assert.doesNotThrow(() => { extractedNoUrl = extractContentFromDOM(createDocument(webHTML), ''); }, 'an empty url must not throw');
+assert.notStrictEqual(extractedNoUrl.title, 'Student records page', 'an empty url is not a student-record page');
+assert.ok(extractedNoUrl.content.includes('Machine Learning'), 'an empty url falls through to generic extraction');
+
+// Test 13: Google Docs text comes from the doc's plain-text export, never the editor
+// DOM, which holds no document text (F7). extractPageContent reads window, document and fetch.
+const docsEditorHTML = `<html><head><title>BIO 200 Syllabus - Google Docs</title></head><body><div id="docs-menubar">File Edit View Insert Format Tools Extensions Help</div><div class="kix-appview-editor">Turn on screen reader support. To enable screen reader support, press Ctrl+Alt+Z</div></body></html>`;
+const docUrl = 'https://docs.google.com/document/u/1/d/1AbC_dEf-GhIjKlMnOpQrStUvWxYz0123456789abcd/edit?tab=t.0';
+globalThis.window = { location: { href: docUrl } };
+globalThis.document = createDocument(docsEditorHTML);
+const docFetches = installFetch(() => new Response('BIO 200 Syllabus\nLearning objectives: analyze enzyme kinetics.', { headers: { 'content-type': 'text/plain; charset=utf-8' } }));
+const exported = await extractPageContent();
+assert.deepStrictEqual(docFetches.map((c) => c.url), ['https://docs.google.com/document/u/1/d/1AbC_dEf-GhIjKlMnOpQrStUvWxYz0123456789abcd/export?format=txt'], 'Docs text must be fetched from the plain-text export');
+assert.strictEqual(docFetches[0].init.credentials, undefined, "the export fetch must keep fetch's default credentials mode");
+assert.ok(exported.content.startsWith('BIO 200 Syllabus'), 'Docs text must come from the export, BOM stripped');
+assert.strictEqual(exported.wordCount, 8);
+assert.strictEqual(exported.emptyReason, undefined);
+
+installFetch(() => new Response('<html></html>', { status: 403, headers: { 'content-type': 'text/html; charset=utf-8' } }));
+const refused = await extractPageContent();
+assert.strictEqual(refused.content, '', 'a refused export must not fall back to the editor DOM');
+assert.ok(refused.emptyReason, 'a refused export must say why nothing was read');
+
+// Signed out, the export answers 200 with an HTML sign-in page.
+installFetch(() => new Response('<html><body>Sign in to continue to Docs</body></html>', { headers: { 'content-type': 'text/html; charset=utf-8' } }));
+const signIn = await extractPageContent();
+assert.strictEqual(signIn.content, '', 'an HTML sign-in page must not be audited as the document');
+assert.ok(signIn.emptyReason, 'an HTML export answer must say why nothing was read');
+
+// Published /d/e/ docs render their text in the DOM and have no export.
+globalThis.window = { location: { href: 'https://docs.google.com/document/d/e/2PACX-1vTabc/pub' } };
+globalThis.document = createDocument('<html><head><title>Syllabus</title></head><body><div id="contents"><p>Week 1 readings.</p></div></body></html>');
+const pubFetches = installFetch(() => { throw new Error('published docs must not be fetched'); });
+const published = await extractPageContent();
+assert.strictEqual(pubFetches.length, 0, 'published docs are read from the DOM, not fetched');
+assert.ok(published.content.includes('Week 1 readings'));
+resetGlobals();
+
+console.log('✅ Task 3 extractor tests passed.');
diff --git a/test/test-manifest.js b/test/test-manifest.js
index 8f7e3bf..9db680d 100644
--- a/test/test-manifest.js
+++ b/test/test-manifest.js
@@ -15,6 +15,25 @@ assert.ok(manifest.permissions.includes('activeTab'));
 assert.ok(manifest.host_permissions && manifest.host_permissions.includes('https://generativelanguage.googleapis.com/*'), 'host_permissions for AI API required');
 assert.ok(manifest.host_permissions && manifest.host_permissions.includes('*://*.instructure.com/*'), 'host_permissions for Canvas API required');
 
+// F4: no declarative content script (its <all_urls> match put "all your data on all
+// websites" on the install prompt). The panel injects the extractor on demand under
+// activeTab; non-instructure Canvas hosts are requested per site at runtime.
+assert.strictEqual(manifest.content_scripts, undefined, 'no declarative content_scripts');
+assert.ok(manifest.permissions.includes('scripting'), 'scripting permission required for on-demand extraction');
+assert.deepStrictEqual(manifest.optional_host_permissions, ['https://*/*'], 'custom-domain Canvas hosts are requested per site');
+assert.ok(!manifest.host_permissions.some((h) => ['<all_urls>', '*://*/*', 'https://*/*', 'http://*/*'].includes(h)),
+  'required host_permissions must stay narrow');
+// chrome.sidePanel.open, called from the icon click, needs Chrome 116.
+assert.ok(Number(manifest.minimum_chrome_version) >= 116, 'minimum_chrome_version must be at least 116 for sidePanel.open');
+
+// The side panel's version badge shows v<major>.<minor> of the manifest version.
+// The badge is hand-written, so a manifest bump alone would leave the panel showing the old version.
+const sidepanelHtml = fs.readFileSync(path.join(__dirname, '../extension/sidepanel/index.html'), 'utf8');
+const badge = sidepanelHtml.match(/<span class="badge-version">([^<]*)<\/span>/);
+assert.ok(badge, 'side panel must show a version badge');
+const [major, minor] = manifest.version.split('.');
+assert.strictEqual(badge[1], `v${major}.${minor}`, `side panel badge ${badge[1]} must match manifest version ${manifest.version}`);
+
 // Storage helper check
 const storagePath = path.join(__dirname, '../extension/shared/storage.js');
 assert.ok(fs.existsSync(storagePath), 'storage.js must exist');
diff --git a/test/test-service-worker.js b/test/test-parser-helper.mjs
similarity index 79%
rename from test/test-service-worker.js
rename to test/test-parser-helper.mjs
index 6d41846..edc7b04 100644
--- a/test/test-service-worker.js
+++ b/test/test-parser-helper.mjs
@@ -1,5 +1,5 @@
-const assert = require('assert');
-const { parseAuditResponse, cleanJsonResponse, getDemoAuditResult, getDemoCourseAuditResult } = require('../extension/background/parser-helper.cjs');
+import assert from 'node:assert';
+import { parseAuditResponse, cleanJsonResponse, getDemoAuditResult, getDemoCourseAuditResult } from '../extension/background/parser-helper.js';
 
 // Test 1: Markdown fenced JSON with '```json'
 const rawLlmResponse = "```json\n{\n  \"summary\": {\n    \"bloomsLevel\": \"Remember\",\n    \"alignmentScore\": \"Moderate\",\n    \"keyTakeaway\": \"Quiz focuses only on memorization.\"\n  },\n  \"findings\": [],\n  \"improvedDraft\": {\n    \"title\": \"Analysis Prompt\",\n    \"content\": \"Compare and contrast\"\n  }\n}\n```";
@@ -28,6 +28,12 @@ assert.strictEqual(parsedWhitespace.summary.bloomsLevel, 'Evaluate');
 assert.strictEqual(typeof cleanJsonResponse, 'function');
 const cleaned = cleanJsonResponse('{"key": "value"}');
 assert.strictEqual(cleaned.key, 'value');
+// Cases from closed bot PRs #99 and #101: arrays, bare whitespace, and malformed JSON.
+assert.deepStrictEqual(cleanJsonResponse('["item1", "item2"]'), ['item1', 'item2']);
+assert.strictEqual(cleanJsonResponse('  \n\t {"key": "value"} \n ').key, 'value');
+assert.throws(() => cleanJsonResponse('```json\n{"key": "value",}\n```'), SyntaxError, 'a trailing comma must not parse');
+assert.throws(() => cleanJsonResponse('{"key": "value"'), SyntaxError, 'truncated JSON must not parse');
+assert.throws(() => cleanJsonResponse('not json'), SyntaxError);
 
 // Test 6: Demo fallback when API key is not configured
 assert.strictEqual(typeof getDemoAuditResult, 'function', 'getDemoAuditResult must be a function');
@@ -42,9 +48,9 @@ assert.ok(demoData.summary.keyTakeaway.includes('Settings'), 'Demo result must n
 assert.ok(demoData.findings.length >= 2, 'Demo result must provide multiple evidence-based findings');
 
 const hasT1 = demoData.findings.some(f => f.tier === 'T1');
-const hasT2 = demoData.findings.some(f => f.tier === 'T2');
+const hasT5 = demoData.findings.some(f => f.tier === 'T5');
 assert.ok(hasT1, 'Demo findings must include T1 evidence tier badge');
-assert.ok(hasT2, 'Demo findings must include T2 evidence tier badge');
+assert.ok(hasT5, 'Demo findings must include T5 evidence tier badge');
 
 assert.ok(demoData.improvedDraft.title, 'Demo draft must include title');
 assert.ok(demoData.improvedDraft.content.includes('Settings'), 'Demo draft must remind user to configure API key in Settings');
diff --git a/test/test-prompts.js b/test/test-prompts.mjs
similarity index 52%
rename from test/test-prompts.js
rename to test/test-prompts.mjs
index e9a5cf9..d63d284 100644
--- a/test/test-prompts.js
+++ b/test/test-prompts.mjs
@@ -1,9 +1,7 @@
-const assert = require('assert');
-const { EVIDENCE_DOMAINS, TIER_METADATA, buildAuditPrompt } = require('../extension/shared/prompts.cjs');
+import assert from 'node:assert';
+import { TIER_METADATA, buildAuditPrompt, buildCourseAuditPrompt } from '../extension/shared/prompts.js';
 
-assert.ok(EVIDENCE_DOMAINS.length >= 10, 'Should include all core idstack research domains');
 assert.ok(TIER_METADATA.T1, 'Tier 1 metadata must exist');
-assert.strictEqual(TIER_METADATA.T1.label, 'Meta-analysis');
 assert.strictEqual(TIER_METADATA.T1.color, '#2f7a4a', 'T1 must match canonical DESIGN.md color');
 assert.strictEqual(TIER_METADATA.T2.color, '#2864a8', 'T2 must match canonical DESIGN.md color');
 assert.strictEqual(TIER_METADATA.T3.color, '#a87726', 'T3 must match canonical DESIGN.md color');
@@ -21,7 +19,6 @@ assert.ok(prompt.includes('Biology 101'), 'Prompt should include content title')
 assert.ok(prompt.includes("Bloom's"), "Prompt should require Bloom's classification");
 assert.ok(prompt.includes('JSON'), 'Prompt should enforce JSON format');
 
-const { buildCourseAuditPrompt } = require('../extension/shared/prompts.cjs');
 assert.strictEqual(typeof buildCourseAuditPrompt, 'function');
 const coursePrompt = buildCourseAuditPrompt({
   title: 'Biology 101',
@@ -35,4 +32,20 @@ assert.ok(coursePrompt.includes('Biology 101'));
 assert.ok(coursePrompt.includes('Constructive Alignment'));
 assert.ok(coursePrompt.includes('courseAudit'));
 
+// A long course: the crawler reads up to 500 assignments, and the prompt keeps
+// a 20,000-char budget for them. It must list whole assignments and say how
+// many it lists, not cut one mid-description under a header that counts all.
+const longCoursePrompt = buildCourseAuditPrompt({
+  title: 'Biology 101',
+  syllabus: 'Objectives: Analyze cellular metabolism.',
+  assignments: Array.from({ length: 80 }, (_, i) => ({ title: `Lab ${i + 1}`, description: 'd'.repeat(600), points: 10 }))
+});
+const assignmentsHeader = /COURSE ASSIGNMENTS & ASSESSMENTS \((\d+) items; (\d+) shown in full below\)/.exec(longCoursePrompt);
+assert.ok(assignmentsHeader, 'course prompt header states the item count and how many are shown in full');
+assert.strictEqual(assignmentsHeader[1], '80', 'the header counts every crawled assignment');
+const listedBlocks = longCoursePrompt.match(/^Assignment \d+: .*\nDescription: .*$/gm) || [];
+assert.strictEqual(listedBlocks.length, Number(assignmentsHeader[2]), 'the header states exactly how many assignments are listed');
+assert.ok(listedBlocks.length > 0 && listedBlocks.length < 80, '80 long assignments exceed the budget, so only some are listed');
+assert.ok(listedBlocks.every((b) => b.endsWith('\nDescription: ' + 'd'.repeat(500))), 'no assignment is cut mid-description');
+
 console.log('✅ Task 2 prompt engine tests passed.');
diff --git a/test/test-sidepanel-logic.js b/test/test-renderer-helper.mjs
similarity index 91%
rename from test/test-sidepanel-logic.js
rename to test/test-renderer-helper.mjs
index ab14f31..eb93548 100644
--- a/test/test-sidepanel-logic.js
+++ b/test/test-renderer-helper.mjs
@@ -1,5 +1,5 @@
-const assert = require('assert');
-const { renderAuditHTML, escapeHtml } = require('../extension/sidepanel/renderer-helper.cjs');
+import assert from 'node:assert';
+import { renderAuditHTML, renderDossierListHTML, escapeHtml } from '../extension/sidepanel/renderer-helper.js';
 
 // Test 1: Full mock audit data rendering
 const mockData = {
@@ -147,7 +147,6 @@ assert.ok(renderedCourse.includes('Course Alignment Matrix'));
 assert.ok(renderedCourse.includes('High (88%)'));
 
 // Test 8: Dossier Item Rendering in Drawer
-const { renderDossierListHTML } = require('../extension/sidepanel/renderer-helper.cjs');
 assert.strictEqual(typeof renderDossierListHTML, 'function', 'renderDossierListHTML must be a function');
 const listHtml = renderDossierListHTML([
   { id: '1', title: 'Week 1 Quiz', pageType: 'Assignment', result: { summary: { bloomsLevel: 'Remember', alignmentScore: 'Low' } } }
@@ -175,7 +174,15 @@ assert.ok(!escapedDossierHtml.includes('<script>'), 'Must escape script tags in
 assert.ok(!escapedDossierHtml.includes('<img src=x>'), 'Must escape img tags in dossier page type');
 assert.ok(escapedDossierHtml.includes('&lt;script&gt;alert(1)&lt;/script&gt;'), 'Must properly escape characters in dossier title');
 
-// Test 9: Consensus Badge and DOI Citation Link Rendering
+// Test 9: malformed model JSON renders instead of throwing (F9). The panel
+// renders inside a runtime.sendMessage callback, so a throw leaves the spinner up.
+assert.ok(renderAuditHTML({ findings: { 0: { tier: 'T1' } } }).includes('Evidence-Based Findings (0)'), 'non-array findings render as none');
+const junk = renderAuditHTML({ findings: [null, 'oops', { tier: 'T5', citation: '[Alignment-1]' }] });
+assert.ok(junk.includes('Evidence-Based Findings (1)'), 'null and non-object findings are dropped');
+assert.ok(junk.includes('[Alignment-1]'), 'a valid finding next to junk still renders');
+assert.ok(renderAuditHTML({ findings: [{ tier: 1, citation: '[CogLoad-1]' }] }).includes('tier-badge tier-1'), 'a non-string tier is stringified');
+
+// Test 10: Consensus Badge and DOI Citation Link Rendering
 const consensusData = {
   summary: {
     bloomsLevel: 'Apply',
@@ -222,4 +229,3 @@ assert.ok(renderedConsensus.includes('rel="noopener noreferrer"'), 'Must include
 console.log('✅ Side panel renderer, escaping, course matrix, dossier & consensus tests passed.');
 
 
-
diff --git a/test/test-service-worker.mjs b/test/test-service-worker.mjs
new file mode 100644
index 0000000..8d1251c
--- /dev/null
+++ b/test/test-service-worker.mjs
@@ -0,0 +1,186 @@
+import assert from 'node:assert';
+import { installChrome, installFetch, jsonResponse, importFresh, resetGlobals, EXTENSION_ID } from './extension-harness.mjs';
+import { saveSettings, getAuditHistory } from '../extension/shared/storage.js';
+
+// A real but short Canvas prompt (13 words): the worker must still audit it.
+const TEXT = 'Post a reflection on chapter four and reply to two peers by Friday.';
+const payload = { url: 'https://canvas.instructure.com/courses/1/assignments/2', title: 'Enzymes Lab', pageType: 'Canvas Assignment', content: TEXT };
+const llmReply = (obj) => jsonResponse({ candidates: [{ content: { parts: [{ text: JSON.stringify(obj) }] } }] });
+const VALID = {
+  summary: { bloomsLevel: 'Apply', alignmentScore: 'Strong', keyTakeaway: 'k' },
+  findings: [{ severity: 'info', tier: 'T1', citation: 'c', observation: 'o', evidence: 'e', recommendation: 'r' }],
+  improvedDraft: { title: 't', content: 'c' }
+};
+
+// One spy for the whole file: it records every AbortSignal.timeout the worker
+// asks for and, when a scenario sets timeouts.fire, returns a signal that has
+// already timed out. Node unrefs the real timer, so nothing waits 25 s.
+const realTimeout = AbortSignal.timeout;
+const timeouts = { requested: [], fire: false };
+AbortSignal.timeout = (ms) => {
+  timeouts.requested.push(ms);
+  return timeouts.fire ? AbortSignal.abort(new DOMException('signal timed out', 'TimeoutError')) : realTimeout.call(AbortSignal, ms);
+};
+
+// Each scenario loads a fresh service worker against a fresh chrome stub. The
+// key goes through saveSettings so the test does not care which storage area holds it.
+async function loadWorker({ apiKey, fetchHandler } = {}) {
+  Object.assign(timeouts, { requested: [], fire: false });
+  const h = installChrome();
+  if (apiKey) await saveSettings({ apiKey });
+  h.fetchCalls = installFetch(fetchHandler || (async (url) => { throw new Error('unexpected fetch ' + url); }));
+  await importFresh('background/service-worker.js');
+  assert.strictEqual(h.listeners.message.length, 1, 'worker registers one runtime.onMessage listener');
+  return h;
+}
+
+// RUN_AUDIT with no API key: demo result, saved to history, no network.
+let h = await loadWorker();
+let res = await h.dispatch({ action: 'RUN_AUDIT', payload });
+assert.ok(res, 'RUN_AUDIT replies (the listener must return true to keep the channel open)');
+assert.strictEqual(res.success, true);
+assert.ok(res.data.summary && Array.isArray(res.data.findings), 'demo audit has the result shape');
+assert.strictEqual(h.fetchCalls.length, 0, 'demo mode makes no network calls');
+assert.strictEqual((await getAuditHistory())[0].title, 'Enzymes Lab');
+
+// RUN_AUDIT with a key: one model call, parsed JSON returned.
+h = await loadWorker({ apiKey: 'k', fetchHandler: async () => llmReply({ summary: { bloomsLevel: 'Apply' }, findings: [] }) });
+res = await h.dispatch({ action: 'RUN_AUDIT', payload });
+assert.strictEqual(res.success, true);
+assert.strictEqual(res.data.summary.bloomsLevel, 'Apply');
+assert.strictEqual(h.fetchCalls.length, 1);
+assert.ok(h.fetchCalls[0].url.startsWith('https://generativelanguage.googleapis.com/'));
+assert.ok(JSON.parse(h.fetchCalls[0].init.body).contents[0].parts[0].text.includes('Enzymes Lab'), 'prompt carries the page');
+
+// Model API failure surfaces as { success: false }, not a hang.
+h = await loadWorker({ apiKey: 'k', fetchHandler: async () => new Response('quota exceeded', { status: 429 }) });
+res = await h.dispatch({ action: 'RUN_AUDIT', payload });
+assert.strictEqual(res.success, false);
+assert.match(res.error, /429/);
+
+// CRAWL_AND_AUDIT_COURSE: crawls the Canvas API, audits, records the course URL.
+// The course has syllabus text, so it is not an empty course.
+h = await loadWorker({ fetchHandler: async (url) => url.includes('include[]=syllabus_body')
+  ? jsonResponse({ name: 'Biology 101', syllabus_body: '<p>Objectives</p>' })
+  : jsonResponse([]) });
+res = await h.dispatch({ action: 'CRAWL_AND_AUDIT_COURSE', payload: { origin: 'https://canvas.instructure.com', courseId: '12' } });
+assert.ok(res, 'CRAWL_AND_AUDIT_COURSE replies (the listener must return true)');
+assert.strictEqual(res.success, true);
+assert.strictEqual(res.courseData.title, 'Biology 101');
+assert.strictEqual((await getAuditHistory())[0].url, 'https://canvas.instructure.com/courses/12');
+
+// Unknown actions get no reply and do not hold the channel open.
+assert.strictEqual(await h.dispatch({ action: 'NOT_A_REAL_ACTION' }), undefined);
+
+// A refused audit replies with an error and is never saved to history.
+async function assertRejected(res, pattern) {
+  assert.strictEqual(res?.success, false, `expected an error, got ${String(JSON.stringify(res)).slice(0, 120)}`);
+  assert.match(res.error, pattern);
+  assert.deepStrictEqual(await getAuditHistory(), [], 'a refused audit must not be saved to history');
+}
+const modelCalls = (h) => h.fetchCalls.filter((c) => new URL(c.url).host === 'generativelanguage.googleapis.com');
+
+// F7: a page with no readable text (restricted tab, PDF viewer) or a lone
+// Modules item title never reaches the model, with or without a key.
+for (const [content, apiKey] of [['', 'k'], ['  \n ', undefined], ['Week 1 Reading', 'k']]) {
+  h = await loadWorker({ apiKey, fetchHandler: async () => llmReply(VALID) });
+  await assertRejected(await h.dispatch({ action: 'RUN_AUDIT', payload: { ...payload, content } }), /too little readable text/);
+  assert.strictEqual(h.fetchCalls.length, 0, `page content ${JSON.stringify(content)} must not be sent to the model`);
+}
+
+// F7: the extractor's or panel's reason for not reading a page is what the user sees.
+const reason = 'idstack does not read student-record pages such as the Gradebook.';
+h = await loadWorker({ apiKey: 'k', fetchHandler: async () => llmReply(VALID) });
+res = await h.dispatch({ action: 'RUN_AUDIT', payload: { ...payload, content: '', emptyReason: reason } });
+await assertRejected(res, /Gradebook/);
+assert.strictEqual(res.error, reason, 'the payload emptyReason is the error shown');
+assert.strictEqual(h.fetchCalls.length, 0);
+
+// F7, course path: no syllabus text and no assignments is an audit of nothing.
+h = await loadWorker({ apiKey: 'k', fetchHandler: async (url) => url.includes('include[]=syllabus_body')
+  ? jsonResponse({ name: 'Sandbox', syllabus_body: '' })
+  : url.includes('/assignments') ? jsonResponse([]) : llmReply(VALID) });
+res = await h.dispatch({ action: 'CRAWL_AND_AUDIT_COURSE', payload: { origin: 'https://canvas.instructure.com', courseId: '12' } });
+await assertRejected(res, /no syllabus text and no assignments/);
+assert.strictEqual(modelCalls(h).length, 0, 'an empty course must not be sent to the model');
+
+// F10 transport: the key travels in a header, not the URL, and the model fetch
+// is bounded below Chrome's 30 s cutoff for an extension fetch.
+h = await loadWorker({ apiKey: 'k', fetchHandler: async () => llmReply(VALID) });
+res = await h.dispatch({ action: 'RUN_AUDIT', payload });
+assert.strictEqual(res.success, true, JSON.stringify(res));
+assert.deepStrictEqual(res.data, VALID);
+assert.deepStrictEqual((await getAuditHistory())[0].result, VALID);
+assert.strictEqual(h.fetchCalls.length, 1);
+const [call] = h.fetchCalls;
+assert.ok(call.url.startsWith('https://generativelanguage.googleapis.com/'));
+assert.ok(!call.url.includes('key='), `API key must not be in the URL: ${call.url}`);
+assert.strictEqual(call.init.headers['x-goog-api-key'], 'k', 'API key travels in the x-goog-api-key header');
+assert.strictEqual(timeouts.requested.length, 1, 'model fetch must be bounded by AbortSignal.timeout');
+assert.ok(timeouts.requested[0] > 0 && timeouts.requested[0] < 30000, `timeout ${timeouts.requested[0]}ms must beat Chrome's 30 s fetch cutoff`);
+assert.ok(call.init.signal instanceof AbortSignal && !call.init.signal.aborted, 'the timeout signal must reach fetch');
+
+// F10: a timed-out model call is explained, not saved. Chrome before 124
+// rejects it with AbortError, so the worker must check the signal, not the name.
+h = await loadWorker({ apiKey: 'k', fetchHandler: async () => { throw new DOMException('The user aborted a request.', 'AbortError'); } });
+timeouts.fire = true;
+await assertRejected(await h.dispatch({ action: 'RUN_AUDIT', payload }), /did not respond within 25 seconds/);
+// Control: any other network error passes through unchanged.
+h = await loadWorker({ apiKey: 'k', fetchHandler: async () => { throw new TypeError('Failed to fetch'); } });
+await assertRejected(await h.dispatch({ action: 'RUN_AUDIT', payload }), /^Failed to fetch$/);
+
+// F9: model output the renderer or dossier compiler cannot handle is refused before it is saved.
+for (const bad of [{ findings: { 0: {} } }, { findings: [null] }, { findings: [{ tier: 1 }] }, { findings: [{ severity: 2 }] }, 'not an object']) {
+  h = await loadWorker({ apiKey: 'k', fetchHandler: async () => llmReply(bad) });
+  await assertRejected(await h.dispatch({ action: 'RUN_AUDIT', payload }), /audit format/);
+}
+
+// F3: only the side panel may start audits. A content script shares sender.id
+// with the panel, so a sender.id check would let any page's injected world
+// trigger credentialed Canvas fetches and write audit history.
+const fakeCanvas = async (url) => url.includes('include[]=syllabus_body')
+  ? jsonResponse({ name: 'Biology 101', syllabus_body: '<p>Course outcomes and weekly schedule</p>' })
+  : jsonResponse([]);
+const contentScript = { id: EXTENSION_ID, url: 'https://evil.example/page', tab: { id: 7 }, frameId: 0 };
+h = await loadWorker({ fetchHandler: fakeCanvas });
+res = await h.dispatch({ action: 'CRAWL_AND_AUDIT_COURSE', payload: { origin: 'https://canvas.instructure.com', courseId: '12' } }, contentScript);
+assert.strictEqual(res, undefined, 'a content-script sender gets no reply');
+res = await h.dispatch({ action: 'RUN_AUDIT', payload }, contentScript);
+assert.strictEqual(res, undefined, 'a content-script sender gets no RUN_AUDIT reply');
+assert.strictEqual(h.fetchCalls.length, 0, 'no credentialed Canvas fetch for a content-script sender');
+assert.deepStrictEqual(await getAuditHistory(), [], 'a content-script sender must not write audit history');
+
+// F3: origin and courseId are spliced into a credentialed fetch URL, so only a
+// bare https origin and a numeric course id are accepted.
+for (const target of [
+  { origin: 'https://victim.instructure.com/api/v1/users/self?x=', courseId: '1' },
+  { origin: 'https://canvas.asu.edu', courseId: '1/users' },
+  { origin: 'http://canvas.asu.edu', courseId: '1' },
+  { origin: 'https://user@canvas.asu.edu', courseId: '1' },
+  {}
+]) {
+  h = await loadWorker({ fetchHandler: fakeCanvas });
+  res = await h.dispatch({ action: 'CRAWL_AND_AUDIT_COURSE', payload: target });
+  assert.strictEqual(res?.success, false, `must reject course target ${JSON.stringify(target)}`);
+  assert.strictEqual(h.fetchCalls.length, 0, `rejected course target ${JSON.stringify(target)} must never be fetched`);
+}
+
+// Positive control: the panel can still crawl a custom-domain Canvas course,
+// and every request stays inside the validated course.
+h = await loadWorker({ fetchHandler: fakeCanvas });
+res = await h.dispatch({ action: 'CRAWL_AND_AUDIT_COURSE', payload: { origin: 'https://canvas.asu.edu', courseId: '42' } });
+assert.strictEqual(res?.success, true, JSON.stringify(res));
+assert.ok(h.fetchCalls.length > 0 && h.fetchCalls.every((c) => c.url.startsWith('https://canvas.asu.edu/api/v1/courses/42')),
+  `every crawl request stays inside the validated course: ${h.fetchCalls.map((c) => c.url).join(', ')}`);
+
+// F4: activeTab is granted only when the icon click reaches action.onClicked;
+// Chrome's openPanelOnActionClick path skips the grant (crbug.com/40904917).
+assert.deepStrictEqual(h.calls.panelBehavior, { openPanelOnActionClick: false },
+  'openPanelOnActionClick must be off: the side-panel toggle path skips the activeTab grant');
+assert.strictEqual(h.listeners.actionClicked.length, 1, 'service worker must handle action.onClicked');
+h.listeners.actionClicked[0]({ id: 3, windowId: 9 });
+assert.deepStrictEqual(h.calls.opened, [{ windowId: 9 }], 'icon click opens the panel synchronously (sidePanel.open needs the gesture)');
+assert.ok(h.sent.some((m) => m.action === 'TAB_ACCESS_GRANTED'), 'icon click tells an open panel to re-read the tab');
+
+resetGlobals();
+console.log('✅ Service worker routing tests passed.');
diff --git a/test/test-sidepanel-state.mjs b/test/test-sidepanel-state.mjs
new file mode 100644
index 0000000..60a404e
--- /dev/null
+++ b/test/test-sidepanel-state.mjs
@@ -0,0 +1,226 @@
+import assert from 'node:assert';
+import { installChrome, installDocument, importFresh, flush, resetGlobals, clickErrors } from './extension-harness.mjs';
+
+// Side panel state across tab switches, errors and clipboard failures (review
+// F8, F14 and the panel side of F9). Each scenario loads the SHIPPED
+// sidepanel.js against the shipped index.html and a stubbed chrome whose
+// worker replies arrive only when the scenario sends them.
+
+const RESULT = {
+  summary: { bloomsLevel: 'Apply', alignmentScore: 'Strong', keyTakeaway: 'k' },
+  findings: [{ severity: 'info', tier: 'T5', citation: '[Alignment-1]', observation: 'o', evidence: 'e', recommendation: 'r' }],
+  improvedDraft: { title: 'Draft', content: 'draft body' }
+};
+const page = (title, url, pageType = 'Canvas Assignment') => ({ url, title, pageType, content: `body of ${title}` });
+const A = page('Assignment A', 'https://x.instructure.com/courses/1/assignments/10');
+const B = page('Assignment B', 'https://x.instructure.com/courses/1/assignments/20');
+const C = page('Assignment C', 'https://x.instructure.com/courses/1/assignments/30');
+const rejectingClipboard = { writeText: () => Promise.reject(new Error('Document is not focused.')) };
+
+// payloads: tabId -> the extractor's result for that tab, or a function
+// returning a promise for it (so the scenario decides when that injection settles).
+async function loadPanel({ tab, payloads, clipboard = { writeText: async () => {} } }) {
+  let currentTab = tab;
+  const pending = []; // sendResponse of each message the worker has not answered yet
+  const document = installDocument();
+  const h = installChrome({
+    tabs: () => [currentTab],
+    onExecuteScript: async ({ target }) => {
+      const p = payloads[target.tabId];
+      return [{ frameId: 0, result: await (typeof p === 'function' ? p() : p) }];
+    },
+    onRuntimeMessage: () => new Promise((resolve) => pending.push(resolve))
+  });
+  // Node 21+ defines globalThis.navigator as a getter; assigning to it throws in a module.
+  Object.defineProperty(globalThis, 'navigator', { value: { clipboard }, configurable: true, writable: true });
+  await importFresh('sidepanel/sidepanel.js');
+  await flush();
+  return {
+    h,
+    document,
+    $: (id) => document.getElementById(id),
+    lastMessage: () => h.sent[h.sent.length - 1],
+    // The worker answers the most recent message.
+    reply: async (response) => { pending.pop()(response); await flush(); },
+    switchTab: async (next) => {
+      currentTab = next;
+      h.listeners.tabActivated.forEach((fn) => fn({ tabId: next.id }));
+      await flush();
+    }
+  };
+}
+
+const scenarios = {
+  async 'result is labelled with the page that was sent (F8)'() {
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A, 2: B } });
+    await p.$('audit-btn').click();
+    await p.switchTab({ id: 2, url: B.url });
+    await p.reply({ success: true, data: RESULT });
+    await p.$('add-to-dossier-btn').click();
+    const [item] = p.h.storage.local.activeDossier;
+    assert.strictEqual(item.title, A.title);
+    assert.strictEqual(item.url, A.url);
+  },
+
+  async 'course audit is labelled from response.courseData (F8)'() {
+    const url = 'https://x.instructure.com/courses/123/modules';
+    const p = await loadPanel({ tab: { id: 1, url }, payloads: { 1: page('Course Modules', url, 'Canvas Modules') } });
+    await p.$('audit-course-btn').click();
+    await flush();
+    await p.reply({ success: true, data: RESULT, courseData: { title: 'BIO 101: Cell Biology' } });
+    await p.$('add-to-dossier-btn').click();
+    const [item] = p.h.storage.local.activeDossier;
+    assert.deepStrictEqual([item.title, item.pageType, item.url],
+      ['BIO 101: Cell Biology', 'Canvas Course (Full Audit)', 'https://x.instructure.com/courses/123']);
+  },
+
+  async 'a slow extraction reply for an old tab is dropped (F8)'() {
+    let releaseB;
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A, 2: () => new Promise((r) => { releaseB = () => r(B); }), 3: C } });
+    await p.switchTab({ id: 2, url: B.url });
+    await p.switchTab({ id: 3, url: C.url });
+    releaseB();
+    await flush();
+    assert.strictEqual(p.$('page-title').textContent, C.title);
+    await p.$('audit-btn').click();
+    assert.strictEqual(p.lastMessage().payload.title, C.title);
+  },
+
+  async 'a slow extraction failure for an old tab does not write its fallback (F8)'() {
+    let failB;
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A, 2: () => new Promise((_, j) => { failB = () => j(new Error('Cannot access contents of the page.')); }), 3: C } });
+    await p.switchTab({ id: 2, url: B.url, title: 'Tab B' });
+    await p.switchTab({ id: 3, url: C.url });
+    failB();
+    await flush();
+    assert.strictEqual(p.$('page-title').textContent, C.title);
+  },
+
+  async 'an error clears the previous result and hides the action bar until the next success (F14)'() {
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A } });
+    await p.$('audit-btn').click();
+    await p.reply({ success: true, data: RESULT });
+    await p.$('re-audit-btn').click();
+    await flush();
+    await p.$('audit-btn').click();
+    await p.reply({ success: false, error: 'AI API error (500)' });
+    const bar = p.document.querySelector('.result-actions-bar');
+    assert.strictEqual(bar.style.display, 'none');
+    await p.$('add-to-dossier-btn').click();
+    assert.strictEqual((p.h.storage.local.activeDossier || []).length, 0, 'Add to Dossier must not save the previous result');
+    await p.$('error-retry-btn').click();
+    await flush();
+    await p.reply({ success: true, data: RESULT });
+    assert.notStrictEqual(bar.style.display, 'none', 'action bar returns on the next success');
+  },
+
+  async 'Retry repeats the action that failed (F14)'() {
+    const url = 'https://x.instructure.com/courses/123';
+    const p = await loadPanel({ tab: { id: 1, url }, payloads: { 1: page('Home', url, 'Canvas LMS Page') } });
+    await p.$('audit-course-btn').click();
+    await flush();
+    await p.reply({ success: false, error: 'Failed to fetch Canvas course info (401)' });
+    await p.$('error-retry-btn').click();
+    await flush();
+    assert.strictEqual(p.h.sent.filter((m) => m.action === 'CRAWL_AND_AUDIT_COURSE').length, 2, 'Retry sends the course crawl again');
+    assert.strictEqual(p.lastMessage().action, 'CRAWL_AND_AUDIT_COURSE');
+  },
+
+  async 'Retry after a switch to a non-course tab does not click the hidden course button (F14)'() {
+    const url = 'https://x.instructure.com/courses/123';
+    const blog = 'https://example.com/blog';
+    const p = await loadPanel({ tab: { id: 1, url }, payloads: { 1: page('Home', url, 'Canvas LMS Page'), 2: page('Blog', blog, 'Web Page') } });
+    await p.$('audit-course-btn').click();
+    await flush();
+    await p.reply({ success: false, error: 'Failed to fetch Canvas course info (401)' });
+    await p.switchTab({ id: 2, url: blog });
+    await p.$('error-retry-btn').click();
+    await flush();
+    assert.deepStrictEqual(p.h.calls.permissionRequests, [{ origins: ['https://x.instructure.com/*'] }],
+      'Retry must not ask for access to the site now active');
+    assert.strictEqual(p.h.sent.filter((m) => m.action === 'CRAWL_AND_AUDIT_COURSE').length, 1);
+  },
+
+  async 'Retry after a switch to another course does not audit that course (F14)'() {
+    const a = 'https://x.instructure.com/courses/123';
+    const b = 'https://x.instructure.com/courses/456';
+    const p = await loadPanel({ tab: { id: 1, url: a }, payloads: { 1: page('Course A', a, 'Canvas LMS Page'), 2: page('Course B', b, 'Canvas LMS Page') } });
+    await p.$('audit-course-btn').click();
+    await flush();
+    await p.reply({ success: false, error: 'Failed to fetch Canvas course info (401)' });
+    await p.switchTab({ id: 2, url: b });
+    assert.notStrictEqual(p.$('audit-course-btn').style.display, 'none', 'course B shows its own course button');
+    await p.$('error-retry-btn').click();
+    await flush();
+    const crawls = p.h.sent.filter((m) => m.action === 'CRAWL_AND_AUDIT_COURSE');
+    assert.strictEqual(crawls.length, 1, 'Retry must not crawl the course now active');
+    assert.strictEqual(crawls[0].payload.courseId, '123');
+  },
+
+  async 'Retry from another page of the same course re-runs that course audit (F14)'() {
+    const root = 'https://x.instructure.com/courses/123';
+    const sub = 'https://x.instructure.com/courses/123/assignments/9';
+    const p = await loadPanel({ tab: { id: 1, url: root }, payloads: { 1: page('Home', root, 'Canvas LMS Page'), 2: page('Lab 9', sub) } });
+    await p.$('audit-course-btn').click();
+    await flush();
+    await p.reply({ success: false, error: 'Failed to fetch Canvas course info (401)' });
+    await p.switchTab({ id: 2, url: sub });
+    await p.$('error-retry-btn').click();
+    await flush();
+    const crawls = p.h.sent.filter((m) => m.action === 'CRAWL_AND_AUDIT_COURSE');
+    assert.strictEqual(crawls.length, 2, 'Retry re-runs the course that failed');
+    assert.strictEqual(crawls[1].payload.courseId, '123');
+  },
+
+  async 'a feedback vote keeps the Audit Another Page button (F14)'() {
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A } });
+    await p.$('audit-btn').click();
+    await p.reply({ success: true, data: RESULT });
+    await p.document.querySelector('.feedback-btn').click();
+    assert.ok(p.$('re-audit-btn'), '#re-audit-btn must survive a vote');
+    assert.strictEqual(p.document.querySelectorAll('.feedback-btn').length, 0);
+    assert.ok(p.document.querySelector('.feedback-row').textContent.includes('Thank you for your feedback!'));
+  },
+
+  async 'Copy to Clipboard reports a rejected write (F14)'() {
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A }, clipboard: rejectingClipboard });
+    await p.$('audit-btn').click();
+    await p.reply({ success: true, data: RESULT });
+    const btn = p.$('copy-improved-btn');
+    await btn.click();
+    await flush();
+    assert.strictEqual(btn.textContent, 'Copy failed');
+  },
+
+  async 'Copy Markdown reports a rejected write without rejecting the handler (F14)'() {
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A }, clipboard: rejectingClipboard });
+    const before = clickErrors.length;
+    const btn = p.$('copy-dossier-md-btn');
+    await btn.click();
+    await flush();
+    assert.strictEqual(clickErrors.length, before, 'click handler must not reject');
+    assert.strictEqual(btn.textContent, 'Copy failed');
+  },
+
+  async 'malformed findings still reach the results state (F9)'() {
+    const p = await loadPanel({ tab: { id: 1, url: A.url }, payloads: { 1: A } });
+    await p.$('audit-btn').click();
+    await p.reply({ success: true, data: { summary: {}, findings: { 0: { tier: 1 } } } });
+    assert.ok(p.$('results-state').classList.contains('active'));
+  }
+};
+
+let failed = 0;
+for (const [name, run] of Object.entries(scenarios)) {
+  try {
+    await run();
+  } catch (err) {
+    failed++;
+    console.error(`not ok - ${name}\n  ${err.message.split('\n').join('\n  ')}`);
+  }
+}
+assert.strictEqual(failed, 0, `${failed} side panel state scenario(s) failed`);
+assert.deepStrictEqual(clickErrors, [], 'a click handler threw or rejected');
+resetGlobals();
+console.log('✅ Side panel labelling, stale refresh, error, retry, vote and clipboard tests passed.');
+process.exit(0); // pending 2 s button-label timers would otherwise hold the process open
diff --git a/test/test-sidepanel.mjs b/test/test-sidepanel.mjs
new file mode 100644
index 0000000..45d429d
--- /dev/null
+++ b/test/test-sidepanel.mjs
@@ -0,0 +1,96 @@
+import assert from 'node:assert';
+import { installChrome, installDocument, importFresh, flush, resetGlobals } from './extension-harness.mjs';
+import { saveSettings } from '../extension/shared/storage.js';
+
+const COURSE_ROOT = 'https://canvas.instructure.com/courses/42';
+const ASSIGNMENT = 'https://canvas.instructure.com/courses/42/assignments/7';
+const CUSTOM_ROOT = 'https://canvas.asu.edu/courses/42';
+
+// Loads a fresh side panel against the shipped index.html and a stubbed chrome
+// whose active tab is `url` and whose extractor injection returns a payload.
+async function openPanel(url, { apiKey, ...chromeOpts } = {}) {
+  const document = installDocument();
+  const h = installChrome({
+    tabs: [{ id: 7, url, title: 'Tab title', active: true }],
+    onExecuteScript: async () => [{ frameId: 0, result: { url, title: 'Extracted title', pageType: 'Canvas Assignment', content: 'Some content' } }],
+    ...chromeOpts
+  });
+  if (apiKey) await saveSettings({ apiKey });
+  await importFresh('sidepanel/sidepanel.js');
+  await flush();
+  return { h, document, $: (id) => document.getElementById(id) };
+}
+
+// Course root: course-audit button shown, tag says so. There is no declarative
+// content script: the panel injects the shipped extractor and reads the
+// script's completion value.
+let p = await openPanel(COURSE_ROOT);
+assert.deepStrictEqual(p.h.calls.executed, [{ target: { tabId: 7 }, files: ['content/extractor.js'] }], 'the panel injects the shipped extractor into the active tab');
+assert.strictEqual(p.$('audit-course-btn').style.display, 'block', 'course audit offered on a course root');
+assert.strictEqual(p.$('page-type-tag').textContent, 'Canvas Course Root');
+assert.strictEqual(p.$('page-title').textContent, 'Extracted title', 'page title comes from the injection result');
+
+// Assignment page: no course-audit button; single-page audit sends the extracted payload.
+p = await openPanel(ASSIGNMENT, {
+  apiKey: 'saved-key',
+  local: { activeDossier: [{ id: 'x', title: 'Saved', result: {} }] },
+  onRuntimeMessage: () => ({ success: true, data: { summary: { bloomsLevel: 'Analyze' }, findings: [] } })
+});
+assert.strictEqual(p.$('audit-course-btn').style.display, 'none');
+assert.strictEqual(p.$('page-type-tag').textContent, 'Canvas Assignment');
+assert.strictEqual(p.$('api-key-input').value, 'saved-key', 'saved key is preloaded');
+assert.strictEqual(p.$('dossier-count').textContent, '1', 'badge shows dossier size');
+p.$('audit-btn').click();
+await flush();
+assert.strictEqual(p.h.sent.length, 1);
+assert.strictEqual(p.h.sent[0].action, 'RUN_AUDIT');
+assert.strictEqual(p.h.sent[0].payload.title, 'Extracted title');
+assert.ok(p.$('results-state').classList.contains('active'), 'results panel shown');
+assert.ok(p.$('results-container').textContent.includes('Analyze'));
+
+// Worker error: rendered escaped in the error card (checked by structure, not entity text).
+p = await openPanel(ASSIGNMENT, { onRuntimeMessage: () => ({ success: false, error: '<b>quota</b>' }) });
+p.$('audit-btn').click();
+await flush();
+const errorMsg = p.document.querySelector('.error-msg');
+assert.ok(errorMsg.textContent.includes('quota'));
+assert.strictEqual(errorMsg.querySelector('b'), null, 'worker error text is escaped, not parsed as HTML');
+
+// F4: the course crawl needs host access to the Canvas site. permissions.request
+// needs the click's user gesture, so it must run before the handler's first
+// await: it has already been called when click() returns.
+p = await openPanel(CUSTOM_ROOT);
+p.$('audit-course-btn').click();
+assert.deepStrictEqual(p.h.calls.permissionRequests, [{ origins: ['https://canvas.asu.edu/*'] }], 'site permission is requested synchronously inside the click');
+await flush();
+const crawl = p.h.sent.find((m) => m.action === 'CRAWL_AND_AUDIT_COURSE');
+assert.ok(crawl, 'with the site permission granted, the course crawl is sent');
+assert.deepStrictEqual(crawl.payload, { origin: 'https://canvas.asu.edu', courseId: '42' });
+
+// Permission denied: no crawl, and the error card says why.
+p = await openPanel(CUSTOM_ROOT, { onPermissionRequest: () => false });
+p.$('audit-course-btn').click();
+await flush();
+assert.ok(!p.h.sent.some((m) => m.action === 'CRAWL_AND_AUDIT_COURSE'), 'no course crawl without the site permission');
+assert.ok(p.document.querySelector('.error-msg').textContent.includes('https://canvas.asu.edu'), 'the error names the site idstack needs');
+
+// The worker announces an icon click (activeTab now covers the tab): re-read it.
+const injectionsBefore = p.h.calls.executed.length;
+await p.h.dispatch({ action: 'TAB_ACCESS_GRANTED' });
+await flush();
+assert.strictEqual(p.h.calls.executed.length, injectionsBefore + 1, 'TAB_ACCESS_GRANTED re-reads the tab');
+
+// A tab switched to without an icon click: Chrome hides its url and title and
+// the injection fails. The header says how to grant access, and an audit sends
+// the reason along so the worker shows it instead of auditing an empty page.
+p = await openPanel(undefined, { tabs: [{ id: 12, active: true }], onExecuteScript: undefined });
+assert.match(p.$('page-title').textContent, /idstack toolbar icon/, 'an unreadable tab tells the user how to grant access');
+p.$('audit-btn').click();
+await flush();
+const unreadable = p.h.sent.find((m) => m.action === 'RUN_AUDIT');
+assert.ok(unreadable, 'the audit request is still sent, so the worker can explain the refusal');
+assert.strictEqual(unreadable.payload.content, '');
+assert.match(unreadable.payload.emptyReason || '', /toolbar icon/, 'an unreadable tab is sent with the reason it could not be read');
+
+resetGlobals();
+console.log('✅ Side panel state tests passed.');
diff --git a/test/test-storage.mjs b/test/test-storage.mjs
new file mode 100644
index 0000000..1209b8c
--- /dev/null
+++ b/test/test-storage.mjs
@@ -0,0 +1,36 @@
+import assert from 'node:assert';
+import { installChrome, resetGlobals } from './extension-harness.mjs';
+import {
+  getSettings, saveSettings, getAuditHistory, saveAuditResult,
+  getDossier, addToDossier, removeFromDossier, clearDossier
+} from '../extension/shared/storage.js';
+
+// Settings round-trip and defaults (independent of storage area, so the key can move)
+installChrome();
+assert.strictEqual((await getSettings()).apiKey, '', 'no key configured by default');
+await saveSettings({ apiKey: 'test-key' });
+assert.strictEqual((await getSettings()).apiKey, 'test-key');
+
+// Audit history: newest first, capped at 20
+for (let i = 0; i < 22; i++) await saveAuditResult({ title: `Audit ${i}` });
+const history = await getAuditHistory();
+assert.strictEqual(history.length, 20, 'history keeps the last 20 audits');
+assert.strictEqual(history[0].title, 'Audit 21', 'newest audit first');
+assert.ok(history[0].timestamp, 'entries are timestamped');
+
+// Dossier: add, replace-by-url, remove, clear
+const a = { url: 'https://canvas.instructure.com/courses/1/assignments/2', title: 'Lab 1', result: {} };
+const b = { url: 'https://canvas.instructure.com/courses/1/assignments/3', title: 'Lab 2', result: {} };
+await addToDossier(a);
+await addToDossier(b);
+await addToDossier({ ...a, title: 'Lab 1 (re-audited)' });
+let dossier = await getDossier();
+assert.strictEqual(dossier.length, 2, 're-auditing the same page replaces its entry');
+assert.strictEqual(dossier[0].title, 'Lab 1 (re-audited)');
+await removeFromDossier(dossier[1].id);
+assert.deepStrictEqual((await getDossier()).map((d) => d.title), ['Lab 1 (re-audited)']);
+await clearDossier();
+assert.deepStrictEqual(await getDossier(), []);
+
+resetGlobals();
+console.log('✅ Storage helper tests passed.');