Give freeform an upper witness without restoring a length cap

[atomchung]

User outcome

A simple lookup stays short. A research-heavy decision can carry several sourced increments when each one changes the call, instead of being implicitly compressed toward a corpus in which every positive exemplar was under 400 characters.

Current behavior / evidence

• Move limits to claim, write, and action boundaries #828 moved missing-input limits from whole-turn refusal to claim/write/action boundaries; its frozen owner blind test preferred the candidate in 3 of 4 scenes.
• feat(challenge): delete the obligation whitelist, keep the seven facts that earn their place (closes #830) #831–Put one canonical exemplar into each surface's generation path (closes #834) #835 then removed obligation recitals and put canonical exemplars onto the generation path. The frozen rerun materially shortened all four target scenes.
• [design·expression] The exemplar corpus is a character cap the contract twice refuses to have #836 found the remaining contradiction: the contract rejects a character cap, but the binding positive corpus demonstrated only the short consequence and almost no multi-source research depth.

Change

• retain the existing compact freeform positives as lower witnesses;
• add one synthetic freeform_research_depth positive with eight distinct increments, explicit source/as-of provenance, an engine-owned portfolio consequence, a contextual research judgment, and observable falsifiers;
• make that scene the one exemplar visible in references/freeform-answers.md;
• state beside it that this is an upper witness, not a minimum/default answer size or permission to narrate irrelevant research.

Scope / non-goals

This is deliberately one surface. It changes no engine, schema, persisted state, transaction lifecycle, execution truth, privacy boundary, or other conversational route. consider, no_book, and weekly_read remain unchanged until owner-live evidence shows they need the same treatment.

Acceptance

• existing compact freeform scenes remain in the corpus;
• freeform stays inside the corpus's 3–5 positive-exemplar range;
• the new scene passes E-5 through E-8;
• the reference/corpus drift gate proves the generation-path copy is canonical;
• product and QA/eval CI pass;
• follow-up owner-live evidence compares one simple lookup and one research-heavy decision before this pattern expands to another surface.

Privacy / rollout / recovery

All issuers, ratings sources, prices, dates, holdings, and deal facts are synthetic. The change is reversible by restoring the previous freeform fence and removing one fixture scene; no migration or recovery path is involved.

Refs #836. This PR does not close it: the issue remains open until owner-live evidence decides whether any other surface needs an upper witness.

[atomchung]

BLOCK — the upper witness repeats the same provenance on two floors.

The three researched claims already satisfy C2 by naming source + date inline (WDGT 8-K, investor presentation, Northstar Ratings). The final caliber line then lists the same three sources and dates again. That teaches the generation path to pay twice for the same provenance, directly against D6 and D7's “nothing appears on two floors” rule—the exact semantic class the deterministic E-5…E-8 checks intentionally cannot catch.

Smallest repair: keep the inline source/date beside each public fact, and reduce the end block to the book/pricing frame plus material gaps, e.g. （帳本 8/14、價格 8/14 收盤；未評反壟斷通過機率與交易後估值）. Apply the same edit to the corpus copy and the fenced generation-path copy, then rerun both CI groups.

Everything else in the two-file slice is appropriately bounded: compact lower witnesses remain, the richer scene is explicitly an upper—not default—witness, the freeform/plain-conversation scope is consistent with the existing route contract, and no engine/state/privacy/execution boundary changes. GitHub cannot accept an approval from the repository owner, so this is recorded as a blocking COMMENT review.

[atomchung]

Blocker repaired and re-reviewed

Commit ff8f11c applies the smallest D6/D7 repair:

• each researched public fact keeps its inline source and date, satisfying C2;
• the end block now carries only the recorded-book/pricing frame and the two material gaps, so the same provenance is no longer paid for on two floors;
• the unrelated wording edit in the pre-existing [bug·voice·consider·P2] A negative post-trade cash balance is owed as two numbers but not as the decision it implies — the answer leads with the boundary instead of "this needs funding or a sale" #778 witness note was reverted.

Final semantic review found no remaining blocker. The PR still changes exactly two synchronized surfaces: the freeform generation-path exemplar and its canonical corpus copy. It changes no engine, schema, persistence, execution state, privacy boundary, or other route.

Verification on the exact final tree:

• focused expression / voice / language / installed-tree / research-prior checks: green;
• local python3 tests/run_all.py --group product: 49/49 suites pass;
• local python3 tests/run_all.py --group qa-eval: 11/11 suites pass;
• GitHub Actions run 33053001530: product-contract and qa-eval-tooling both pass on Python 3.11 and 3.12; network-smoke is skipped by PR policy.

Mechanical and semantic acceptance for this PR are complete. Behavioral acceptance is deliberately still open: #836 requires the frozen simple-lookup versus research-heavy owner-live comparison before this pattern expands to another surface. Merge remains an owner decision.