Fix current task startup guidance while preserving frozen inputs

docs/research/agent-procedure.md

@@ -118,16 +118,39 @@ It reads [the task catalogue](agent-task-catalog.json), draws a
 cards with independently selected methods. It reports your checkout's commit
 and selected-record hash. It does not contact GitHub, reserve work, or establish
 that a task is open. A deterministic seed is a reproducibility aid, not novelty.
-Each card supplies a first working command, expected evidence, a next experiment,
+Each card supplies a first-step command, expected evidence, a next experiment,
 its connection to the bottleneck, allowed variations and scope-change triggers.
 
+**Resolve the generic orientation command before running a historical card.**
+If its `first_step.command` is `python3 -B scripts/research_baseline.py`, use the
+current orientation command in [current priorities](current-priorities.md).
+For the selected #4470 productive-helper schema, that command is:
+
+```sh
+python3 -B scripts/research_selected_baseline.py
+```
+
+The older helper expects `bit_lower` and `complex_lower`, which this certificate
+does not contain. This substitution applies only to the generic selected-baseline
+step; retain the card's question, inspection instructions and any named scientific
+replay command. A future selection must be checked against its own reviewed schema.
+The [startup compatibility audit](task-startup-compatibility.md) explains why the
+historical catalogue, picker and baseline bytes remain intact. The replacement
+authenticates selected sources and checks metadata consistency; it is not a cold
+replay or supplier admission.
+
 In a sparse checkout, the helper can validate omitted catalogue references when
 they are regular files in `HEAD` with matching skip-worktree index entries. It
 reports these as `sparse_checkout.omitted_catalogue_paths`. For absent paths,
 ordinary deletions and staged content replacements still fail. Materialize the
 chosen card's inputs before running its commands. This reference check does not
 resolve replay dependencies.
 
+For missing selected-baseline inputs, run
+`python3 -B scripts/research_baseline_sparse.py`, inspect its materialization
+command, and use its schema-aware `next` command after materializing the inputs.
+Its omission report does not authenticate sources; the baseline still must pass.
+
 Use the overlap helper before choosing a card and refresh before submission:
 
 ```sh

docs/research/task-startup-compatibility.md

@@ -0,0 +1,130 @@
+# Task startup and frozen-input compatibility
+
+At canonical main `624de9dc66858af0945e5f8a0942488d62280d43` (priority revision 7),
+16 of 22 task cards emit `python3 -B scripts/research_baseline.py` as their first
+step. The catalogue's `selection_context.orientation_command` and current
+priorities already name `research_selected_baseline.py`; the picker does not
+include `selection_context` in its output. Following a card alone therefore
+reaches the historical certificate schema.
+
+Use the [standing procedure's generic-command substitution](agent-procedure.md#1-orient-briefly-then-choose-a-distinct-task).
+It routes current contributors to the existing source-bound successor while
+leaving frozen cards, helpers and scientific commands unchanged. This is a
+navigation repair and scoped dependency audit, not a new certificate adapter,
+research priority, supplier or kappa claim.
+
+## Exact dependency boundary
+
+The inventory below covers every tracked top-level `research/*/SOURCES.json`
+and `research/*/VERIFY.json` at that base: **5,854 files**, each checked against
+its Git blob. Counts refer to exact repository paths in standard manifest fields.
+Nested vendor manifests, code imports, archives and undeclared reads are outside
+this audit. A provenance pin reads historical Git bytes; a live input pin checks
+working bytes. A declared replay input selects a package when the path changes,
+even if its source pin is only provenance.
+
+| Path | Live `input` rows | `provenance` rows | Declaring `VERIFY.json` files |
+| --- | ---: | ---: | ---: |
+| `docs/research/agent-task-catalog.json` | 16 | 2,676 | 57 |
+| `scripts/choose_research_task.py` | 1 | 8 | 1 |
+| `scripts/research_baseline.py` | 6 | 4 | 7 |
+| `docs/research/agent-procedure.md` | 0 | 3 | 0 |
+
+The catalogue's 16 live pins already differ from the catalogue bytes at this
+base. Their existence is **not** evidence that this proposed change newly breaks
+16 passing packages. Editing the catalogue would nevertheless select 57 declared
+dependents, requiring a separate compatibility decision. The procedure has only
+historical provenance pins in this inventory, so this repair changes that active
+guide and adds this note. No source receipt is rewritten or silently repinned.
+
+For example, the catalogue is a live input in
+[`clean-selector-embedding/SOURCES.json`](../../research/clean-selector-embedding/SOURCES.json),
+whereas the procedure is provenance in
+[`p104-newton-interfaces/SOURCES.json`](../../research/p104-newton-interfaces/SOURCES.json).
+The substring reference in `kappa-rebuild/SOURCES.json` names its separate
+`provenance/public-final/docs/research/agent-procedure.md` copy, not the live guide.
+
+## Reproduce the inventory
+
+From a checkout containing the pinned base, run this Python snippet. It reads
+committed objects, so sparse files need not be materialized; a partial clone may
+fetch the required blobs. It does not execute research packages. Git replacement
+objects are disabled for these reads.
+
+```python
+import io, json, subprocess
+
+BASE = "624de9dc66858af0945e5f8a0942488d62280d43"
+TARGETS = ["docs/research/agent-task-catalog.json",
+           "scripts/choose_research_task.py", "scripts/research_baseline.py",
+           "docs/research/agent-procedure.md"]
+def git(*args, **kwargs):
+    return subprocess.check_output(["git", "--no-replace-objects", *args], **kwargs)
+
+entries = []
+for entry in git("ls-tree", "-r", "-z", BASE).split(b"\0"):
+    if not entry:
+        continue
+    metadata, name = entry.split(b"\t", 1)
+    parts = name.decode("utf-8").split("/")
+    if (len(parts) == 3 and parts[0] == "research"
+            and parts[2] in {"SOURCES.json", "VERIFY.json"}):
+        entries.append((name.decode("utf-8"), metadata.split()[2]))
+stream = io.BytesIO(git("cat-file", "--batch",
+                       input=b"".join(oid + b"\n" for _, oid in entries)))
+counts = {p: {"input": 0, "provenance": 0, "declared": 0} for p in TARGETS}
+for name, oid in entries:
+    actual, kind, size = stream.readline().split()
+    if actual != oid or kind != b"blob":
+        raise ValueError("Unexpected Git object: " + name)
+    data = json.loads(stream.read(int(size)))
+    if stream.read(1) != b"\n":
+        raise ValueError("Invalid Git batch framing")
+    if not isinstance(data, dict):
+        continue
+    for target, result in counts.items():
+        if name.endswith("/VERIFY.json"):
+            result["declared"] += target in data.get("inputs", [])
+        elif isinstance(data.get("files"), list):
+            for row in data["files"]:
+                if isinstance(row, dict) and row.get("path") == target:
+                    usage = row.get("usage")
+                    if usage in {"input", "provenance"}:
+                        result[usage] += 1
+print(json.dumps({"base": BASE, "manifests": len(entries), "paths": counts}, indent=2))
+```
+
+## Evidence and remaining work
+
+The selected #4470 closure passes the existing source checker. On those same
+authenticated inputs, the historical command fails with `KeyError: 'bit_lower'`;
+the successor passes in ordinary and optimized Python with identical reports.
+The result remains `PASS_RESEARCH_BASELINE_BINDINGS_ONLY`. Existing selected
+baseline controls pass 13 tests, and assignment controls pass 16 tests on native
+Windows / CPython 3.12.0. All eight coordination tests pass with `-X utf8`; the
+default cp1252 run has one existing `read_text()` decoding error on
+`autoresearch.md`. The exact inventory snippet reproduces the table from Git
+objects, independently of the working-file/blob comparison. The registered
+package planner selects zero packages for the two changed documentation paths
+and 57 for a hypothetical catalogue edit. No mathematical cold replay, broad
+historical replay or Lean build was run.
+
+**Open next question:** migrate the 16 emitted generic commands after an explicit
+decision on the 57 declared dependents. Start with the exact-path inventory above,
+distinguish existing drift from a new incompatibility, and preserve each package's
+original task and source history. This matters because callers of the picker that
+skip the standing procedure will still see the historical command. Hidden runtime
+dependencies remain a blocker to claiming a complete compatibility migration.
+
+This refines `scope-navigation`, selected with nonce
+`cc0ba7957f5aa29a9b8e5a07a7a6ec93`, and continues the migration handoff in
+[Discussion #4458](https://github.com/CrocSwap/integer-mult-bounds/discussions/4458#discussioncomment-19134442).
+Credit Douglas Colkitt and retained onboarding/source-checker contributors;
+Mihir Srivastava / [#4463](https://github.com/CrocSwap/integer-mult-bounds/pull/4463)
+for the productive-helper adapter; lee-phys /
+[#4631](https://github.com/CrocSwap/integer-mult-bounds/pull/4631) for schema-aware
+sparse diagnosis; and Mister-Meeseeks / the credited Claude agent in
+[#3725](https://github.com/CrocSwap/integer-mult-bounds/pull/3725) for the
+immutable-helper precedent. Scientific sources retain the selected contributor
+map's attribution. Prepared by Mihir Srivastava with substantial OpenAI Codex
+assistance; no subagents. Apache-2.0.