feat(decision-framing): add bounded research priors

[atomchung]

Scope

• Adds four audited, bounded research priors for the no-book route.
• Adds the single SKILL.md → decision-framing.md → research-priors.md loading path.
• Defines research-aware strategy framing as baseline → applicable strategy map → zero or one discriminating question last.
• Keeps ordinary no-book single-trade framing separately bounded at three questions.

Boundaries

• No engine arithmetic, schema, state, persistence, runtime web client, fund recommendation, allocation, suitability conclusion, forecast, or trade verdict.
• No Generic Parity runner, case/witness bank, receipt platform, judge plumbing, or evaluator documentation.

Validation

• python3.12 tests/test_research_priors.py
• python3.12 tests/test_doc_language.py
• python3.12 tests/test_output_voice.py
• python3.12 tests/run_all.py (full offline suite)
• Independent review, including a re-review after two scope/order corrections.

Evidence status

This is deterministic implementation evidence only. Owner-live, cross-model Generic Parity, and Antigravity testing are explicitly deferred.

[atomchung]

[BLOCK] Independent review found a private-boundary escape: --output-dir is caller-controlled and the runner writes exact raw model outputs under it without rejecting the effective coach-state root. python3 evals/run_generic_parity.py --output-dir ~/.trade-coach/parity would place evaluator output inside coach state, contradicting #715's receipt boundary. Please normalize and reject equal/nested output paths (including symlink resolution), add a regression test, then rerun review. Python 3.11 and 3.12 CI are green, but this remains a merge blocker.

[atomchung]

[BLOCK] args.output_dir is not checked against the effective coach-state root before creating a directory and later writing A*.raw.txt and receipt.json. This permits raw evaluator output inside ~/.trade-coach when supplied by the caller, violating the stated outside-coach-state receipt boundary. Reject equal/nested paths after resolving symlinks (and test it).

[atomchung]

Review — blocked; do not merge or mark ready yet

Reviewed draft head 8e2ba87d3c81086554253ffbb3da6d37982af960 against #715 owner comment 5151689090, #27, and the already-open #726.

P1 — this creates a second owner for the same #715 product contract

#715 explicitly authorized one small draft/reviewable PR for the A01 loop. PR #726 already changes references/decision-framing.md, adds A01/A07/A10 witnesses, and owns the same visible outcome. This PR changes the same authority again while also adding a second evaluation implementation. Keeping both open makes it unclear which branch owns the product wording and guarantees drift/conflict.

Pick one surviving product PR. My recommendation is to keep the smaller #726 product leaf, rebase it onto main@5ed29d9, and treat any probe runner as local iteration tooling or a separately owned follow-up only if evidence proves it must be committed. Do not merge both.

P1 — the route order contradicts the accepted A01 outcome

decision-framing.md says the order is baseline → exception gate → strategy map → at most one question, but then says: “First ask the material exception: is the money needed…”. That restores question-first intake before the user receives the complete strategy-class map and can also create a second question after the map. The owner disposition requires available value and the bounded map first, followed by at most one discriminating question.

State liquidity/narrow-exposure as a material limitation/branch condition in the answer. Do not ask it before the map. Add a deterministic witness/mutation that fails if any question precedes the strategy-class completion.

P1 — the evaluator can report candidate_pass with no semantic evidence

generic_parity.receipt_disposition([], {}) returns candidate_pass: the classification list is empty, so neither ambiguity nor disagreement fires. This is a fail-open release-evidence bug. An empty report, a missing axis, or a partial axis set must be ambiguous/failure, never a pass. Add tests for empty, missing, and unknown classifications, and require the complete AXES set before candidate_pass.

P2 — the PR exceeds the authorized first loop

The owner allowed one A01 product correction and one A01/A07/A10 rerun; it explicitly did not authorize the full #716 catalogue or broad evaluator work. This PR adds four shared research priors, 11 changed files, a new case/witness bank, a runner, receipt format, docs, and judge plumbing. Those may be useful, but they are separate durable owners from the smallest #715 product cut and should not be bundled merely because quota prevented the live rerun.

Evidence status

Standard CI is green on Python 3.11/3.12, which proves the branch is mechanically stable. The required live evidence is still ambiguous due provider quota, and CI does not resolve the ownership/order/fail-open findings above.

Required before re-review:

1. resolve feat(no-book): map strategy class before exit #726/feat(decision-framing): add bounded research priors #727 into one owning product PR;
2. preserve value/map-before-question ordering;
3. make semantic receipt aggregation fail closed on absent/partial axes;
4. narrow or re-home the [research·product] Cross-route evidence-backed decision priors — 建立不低於通用模型的回答基準 #716 research authority and evaluator expansion;
5. rerun the frozen A01/A07/A10 probe when independent-family capacity is available.

#713 remains the merge blocker; no owner-live acceptance is supplied by this PR.

[atomchung]

Owner disposition — surviving PR, but narrow to the product/reference slice

The owner now explicitly wants the next test candidate to have basic research-aware no-book answers. This supersedes the earlier recommendation to keep #726: PR #727 becomes the one surviving implementation owner, provided it is narrowed before re-review.

Required changes tonight:

1. Keep the product slice only

   • retain references/research-priors.md with at most the four current priors;
   • retain the thin decision-framing.md consumption path;
   • add/verify the smallest guaranteed route pointer needed for the real no-book host to load the catalogue;
   • retain only deterministic witnesses needed to prove the product contract.

2. Remove the evaluator expansion from this PR

   • remove evals/run_generic_parity.py, evals/generic_parity.py, new case/witness banks, judge plumbing and evaluator-design documentation unless an existing repository test genuinely requires a tiny reusable helper;
   • no new receipt format or committed cross-model harness in this slice;
   • tomorrow's live A01/A07/A10 probe may use existing/local tooling and can motivate a separate owner only if needed.

3. Fix the visible interaction order

   • state the research-backed baseline and applicable limitation/exception conditions;
   • provide the bounded strategy-class map;
   • only then ask at most one discriminating question;
   • do not ask the liquidity/material-exception question before the user receives the available baseline and choices.

4. Make the route actually consume the prior

   • a new reference file alone is dead documentation;
   • trace the supported no-book entry from SKILL.md/route authority to decision-framing.md and the new catalogue, and add a structural regression that fails if the pointer or catalogue disappears;
   • do not duplicate the full priors across prompt surfaces.

5. Preserve strict boundaries

   • no named product/fund recommendation, ranking, geography choice, allocation percentage, cash-reserve amount, moving average, stop percentage, timing window, expected return, suitability claim, persistence, engine/schema/domain object or runtime web client;
   • thematic/sector ETF is not broad diversification;
   • tactical governance cannot leak into a long-horizon market policy;
   • recurring future savings remains distinct from staging already-available cash.

No owner-live or billable/cross-model fast probe tonight. Run focused deterministic tests, full offline suite and standard Python 3.11/3.12 CI. Update #716/#715 with exact head and evidence, then stop at ready-for-independent-review. Do not merge without a separate owner merge authorization.

[atomchung]

Independent re-review of c5b79a31bb77c1d96ef620bd931e05394fc7fac7: no findings.

The previous Generic Parity thread is outdated because its runner and raw-output path were removed from this PR. The current diff contains only the bounded reference slice; it keeps the baseline/map-before-question order, caps the research-aware path at one final question, and has deterministic loading/order mutation coverage. Full offline tests passed locally on Python 3.12; GitHub Actions passed Python 3.11 and 3.12 (run 30706651485). No owner-live, cross-model, or Antigravity acceptance is claimed.