feat(review): the entry-thesis question quotes what the user already said, instead of asking them to remember it (refs #636)

[atomchung]

What the user sees

A first review asks every un-thesised holding what its entry thesis was, and offers five buckets. For a position the user had already discussed through consider, their exact words were sitting in trade_evaluations.jsonl the whole time — stored verbatim, read by nothing outside consider itself.

before  You are holding AAA at a cost basis of about $5,000.
        When you first entered this position, what was your thesis?

after   You are holding AAA at a cost basis of about $5,000.
        On 2025-12-28 you said: “…” Was that the thesis you entered on?

Same question kind, same choices, same event, same slot in the #291 density band. The question is asked better, not asked twice.

Why recall does not route through _evaluation_reconciliation

That surface looked like the obvious source and is the wrong one, twice over:

• it is open-evaluations-only, and a resolved evaluation is still something the user said — an entry the user later marked acted is exactly the case worth recalling;
• its schema declares it "A supply-side fact surface, not a card section or a question", so consuming it as a question source would contradict its own contract, and its item shape is additionalProperties: false.

_evaluation_recall reads the store directly and does not read decision at all.

Boundaries, each locked by a test

rule	why
a statement recorded after the cycle's entry is never quoted	it describes a different decision than the one this question asks about
the latest qualifying statement wins	an earlier one it superseded is not the user's account of why they entered
an evaluation with no reason/why_now is dropped, not quoted empty	consider runs fine with no --decision-context; a recalled blank is worse than the canned question it replaces
a cycle whose start date is unknown (TICKER#unknown) recalls nothing	a statement must not attach to a cycle whose entry nobody knows

Same-day counts: contemplation and fill routinely land on one date.

This states a fact and never a cause. The stem says the user said this, on this date. It never says the position was opened because of it — the same boundary _evaluation_reconciliation documents for itself.

What this deliberately does not do

• initial_thesis stays in KNOWN_UNWIRED. This improves the question; the answer still has no consumer. [design·P1] Two questions collect an answer nothing reads: initial_thesis and exit_consistency (proved by a byte-identical card) #429 is not fixed here, and claiming it was would be false.
• No new free-text field. The note slot already exists and is already stored — only the invitation is missing. Wiring it before a reader exists would make [design·P1] Two questions collect an answer nothing reads: initial_thesis and exit_consistency (proved by a byte-identical card) #429 worse, not better (see [epic·design] How we ask, how we collect, how we distill — one view over the question layer #398).
• Nothing for a user with no book. consider requires one, so cold start is untouched. That path needs the thesis aggregate refactor and three product-contract changes, tracked on [feat·M2] DecisionIntent memory — automatically record a live trade intent, then resolve it from execution evidence or one follow-up #636.

Verification

• python3 tests/run_all.py — 49/49 suites pass, re-run after rebasing onto main@22417a7.
• Two mutations, both caught, re-verified on the rebased tree: forcing _evaluation_recall to return empty → red; removing the entry-date boundary → red.
• New tests confirmed to actually execute, not skip.

🤖 Generated with Claude Code