feat(qa): synthetic consider user walkthrough

[atomchung]

1. LLM role = synthetic user, never final-answer generator — qa/synthetic_user.py and tests/test_synthetic_walk.py constrain it to the current allowed action.
2. Execute the real consider-ai-momentum walkthrough in isolated FOMO QA state — qa/run_synthetic_walk.py invokes review.py consider with isolated HOME and TRADE_COACH_HOME.
3. Capture the exact product-visible answer byte-for-byte — skills/fomo-kernel/engine/consider_surface.py, review.py --product-surface, and the capture equality assertion.
4. Build the deterministic schema-v3 artifact and call the existing judge — qa/candidate_builder.py, evals/judge_trade_answers.py --answer-file --source-fixture, and the existing eval: add first TradeEvaluation LLM judge slice #705 candidate validators.
5. Stop after one scenario, one persona, and the consider route — tests/test_synthetic_walk.py pins consider-ai-momentum / ai-momentum-investor / consider.

Scope

• One command defaults to a real agy synthetic user, runs the actual route, captures its final answer, emits the candidate artifact, calls the existing judge, and writes a combined receipt with workflow, deterministic, semantic_judge, and owner_acceptance=owner_unreviewed.
• A live agy walk completed the real context queue, final product capture, and UX receipt using only isolated synthetic data. Its receipt has workflow=pass, deterministic=pass, and semantic_judge=skipped because that verification intentionally used --no-semantic-judge.
• The offline test lane uses --user-backend stub --no-semantic-judge; the judge plan has been verified against the captured artifact. The runner records semantic_judge=unavailable if the existing judge backend cannot return, rather than dropping the combined receipt.
• No second scenario, dashboard, benchmark, persona library, judge rubric, eval framework, README, or documentation expansion.

Verification

• PYTHONPYCACHEPREFIX=/tmp/fomo-pycache python3 tests/run_all.py
• python3 qa/run_synthetic_walk.py consider-ai-momentum --user-backend stub --no-semantic-judge --output-dir /tmp/fomo-synthetic-issue718
• python3 evals/judge_trade_answers.py --answer-file /tmp/fomo-synthetic-issue718/candidate.json --source-fixture /tmp/fomo-synthetic-issue718/source-fixture.json --plan
• PYTHONPYCACHEPREFIX=/tmp/fomo-pycache python3 qa/run_synthetic_walk.py consider-ai-momentum --user-backend agy --no-semantic-judge --output-dir /tmp/fomo-synthetic-live-718-final

Refs #718

[atomchung]

Post-merge review finding — resolve before freezing #713

PR #723's isolated QA orchestration is useful and its Python 3.11/3.12 CI is green, but the merged cut also introduced a generic runtime/product surface that does not satisfy the contract it claims to capture.

P1 — --product-surface is scenario-specific copy exposed as a generic consider answer

skills/fomo-kernel/engine/consider_surface.py::render() unconditionally opens the case with:

• “If the demand observation proves durable…”
• “…changing the cap can be a deliberate choice.”

Those statements are not derived from the supplied premise, side, context or challenge. The public CLI flag is generic (review.py consider --product-surface) and accepts sell decisions, non-demand reasons, and cases with no cap-shaped tension. The focused production test covers only the one AI-momentum buy fixture, so this remains green while another valid consider scene receives a fabricated decision frame.

The same renderer is English-only and emits raw internal topic labels (The frozen {topic} fact...), so it is not the existing localized/output-voice product surface either.

Contract consequence

Issue #718 says this lane is additive maintainer tooling, not a runtime feature, and that product runtime behavior stays unchanged. It also requires capturing the exact existing product-visible answer byte-for-byte. The current implementation instead creates a new deterministic answer renderer inside the engine and then captures that new surface. Therefore either:

1. this is a QA-only fixture renderer, in which case it must not be exposed or described as the generic product answer and cannot prove exact production-answer capture; or
2. this is a new product answer path, in which case it needs a separate product owner, generic scenario/locale/output-voice coverage, and acceptance evidence beyond [qa] Add an LLM-driven synthetic FOMO QA walkthrough #718.

Smallest safe correction

Do not widen this into a response framework. Before #713 freezes a candidate:

• remove the generic product claim/runtime exposure or constrain the adapter so it cannot render outside the declared synthetic scenario;
• ensure the synthetic walk captures the actual product answer path rather than a scenario-specific substitute;
• add a regression using at least one non-demand/sell consider scene that fails on the current hard-coded text;
• preserve the isolated synthetic-user, exact-byte receipt, schema-v3 builder and existing judge work that is genuinely QA-only.

Until this boundary is resolved, main@3eb12f7 should not be named the #713 reliability candidate. Green CI proves the implemented behavior is stable, not that this is the correct product surface.