fix(card): a returning user's standing rule is the one they committed to, not this period's prescription (closes #645)

[atomchung]

Closes #645. Implements the owner's 2026-07-31 handoff comment (option 2), step for step.

Does the persisted commitment carry a comparable dim? Yes — and it needed checking

The issue asked this before option 2 could be costed. It does, on every commitment that has a dimension at all, and nothing new had to be stored:

• skills/fomo-kernel/schemas/answers.schema.json — commitment declares dim (optional, alongside rule/metric_key/goal).
• skills/fomo-kernel/engine/review.py:3197 — _candidate_rules builds every candidate row as {"id": ..., "dim": dim_id, "rule": ..., "metric_key": ..., "goal": "down"}, where dim_id = card_renderer.dimension_id(dim) (line 3182). The canonical id, not the legacy label.
• skills/fomo-kernel/engine/review.py:5027 — _resolve_commitment takes chosen = dict(candidates[selected]) for a candidate choice, so dim survives verbatim into the stored commitment; line 5031 carries the agent-supplied choice.get("dim") for a custom one.
• skills/fomo-kernel/engine/review.py:3906 — state_snapshot.prior_commitment is _with_condition_line((previous or {}).get("commitment")), i.e. that same stored object, so the renderer reads it frozen from the plan and never reconstructs it.
• skills/fomo-kernel/engine/trade_recap.py:2191 — state["rule_dim"] = actionable[0].get("dim") is the legacy label (部位 sizing), since prescribe writes those.

So the two live in two namespaces, and card_renderer.dimension_id (card_renderer.py:260) is the existing, deterministic normalizer between them — the same one _candidate_rules already calls. Both sides go through it: dimension_id(prior["dim"]) == dimension_id(state["rule_dim"]).

The task's warning about session.PKEY holds and is why nothing else was used: PKEY (session.py:36) maps metric_key into a problem namespace (oversize, avgdown_breach, concentration, …) which is not the rules copy namespace rule_dim resolves in, and concentration has three metric_keys feeding it, so it is not invertible. No PKEY→dim table was added.

Which reading I implemented, and why

Owner step 2 — rule_standing only on proven agreement — and step 1 — read the commitment from the frozen snapshot, never reconstruct it — together mean the quoted rule has to change too, not only the wrapper choice. Quoting localized_rule(state["rule_dim"]) inside the continuity wrapper is the reported defect; leaving it there for the agree case would just narrow the same defect to "agrees on dimension, still not the sentence the user wrote" (a custom rule on averaging_down would still be replaced by the catalog's averaging_down text). So the standing rule is now prior_commitment["rule"], quoted verbatim — the same treatment _reconciliation_lines already gives it in the card's opening (card_renderer.py:2347), which also makes the two places a card names the standing rule agree with each other for the first time.

Four arms, in card_renderer._next_block:

Prior commitment	Dims	Wrapper	Quotes
none persisted	—	rule_pending (#546, byte-stable)	this period's prescription
persisted, no comparable dim	—	rule_pending (fail-safe)	this period's prescription
persisted	agree	rule_standing	the committed rule
persisted	differ	rule_diverged (new)	committed rule and prescription

Owner step 4 (fail safe rather than guess) applies to the second row: _slot_commitment (review.py:4953) drops dim from a condition-slot commitment deliberately, and a custom commitment need not carry one. rule_pending is the one wrapper in this branch that asserts neither continuity nor divergence, so the claim is suppressed by falling back onto it — no new memory object, no new state field, no new key for that case. The user does not lose the fact: _reconciliation_lines still opens the card with their standing rule and its then/now.

A missing state["rule_dim"] still falls through to the generic block_missing.rule line exactly as before (the interpolation guard became all(fields.values())).

The exact copy, all three locales

One new key, block_missing.rule_diverged. Each phrased natively, not translated word-for-word; rule_standing / rule_pending are untouched.

• en — The standing rule is still "{standing}"; this period's leading issue is a different one, so "{recommendation}" stays a recommendation until you choose it.
• zh-TW — 現行規矩仍是「{standing}」；這期最該處理的是另一件事，「{recommendation}」在你選定之前都只是建議。
• zh-CN — 现行规矩仍是“{standing}”；这期最该处理的是另一件事，“{recommendation}”在你选定之前都只是建议。

It replaces the continuity sentence rather than adding one — one paragraph for one paragraph, so net card line count is unchanged and the standing no-explanatory-Note ruling is respected. Both renderers share it through _card_structure; Markdown's compact text-first scan (_read_first_panels) quotes this same panel, so there is no second mirror. docs/output-contract.md's Block-4 row moved with it, plus a dated register row in the same register as #546's.

Mutations run, and what reddened

Restored with cp from a saved copy between runs (never git checkout), __pycache__ cleared each time. All five red; m1 is the owner's named target — reverting the quoted rule to state.rule_dim on a fixture whose prior commitment intentionally differs.

Mutation	Result
m1 — revert to pre-#645 main (one wrapper decision, always quoting state.rule_dim)	RED: the standing rule must be the one the user committed to, quoted verbatim
m2 — keep the divergence line but fill its standing slot from the prescription	RED: same assertion
m3 — agree arm quotes the prescription instead of the committed rule	RED: the continuity line quotes the committed rule, not the prescription that happens to share its dimension
m4 — drop the fail-safe, assert divergence on an uncomparable commitment	RED: must not be presented as diverging either
m5 — drop the fail-safe the other way, assert continuity	RED: must not be presented as unchanged

The test is tests/test_card_html.py::test_a_returning_users_standing_rule_is_the_one_they_committed_to — diverge / agree / uncomparable, Markdown and HTML, zh-TW and en, plus a direct three-locale assertion that rule_diverged exists with both placeholders in zh-CN too. It is registered in that file's explicit runner list (confirmed running, not silently absent: the suite goes 62 → 63 tests, and it reddens on demand above). Its banned-sentence expectations are derived from the wrapper catalog rather than pinned as literals, so the ban cannot drift away from the sentence it bans.

#546's own test keeps its meaning: its returning-user fixture now carries dim: "position_sizing" to match rule_dim, which is what the narrowed contract requires for continuity.

Persona sweep

python3 tests/persona_sweep.py --baseline <origin/main checkout>/skills/fomo-kernel/engine
persona sweep: 156 bundles x 3 surfaces — PASS: all gates pass
NOTE  html vs baseline: 156/156 byte-identical

Zero drift, Markdown and HTML (Markdown byte-parity is a gate, not a note, so PASS is the statement). Classifying every swept bundle against the new branch explains why, and it is the useful half:

• 8 .second.preview bundles reach the agreement arm (ai_holder, momentum, panic_seller, pyramid × 2 locales), all position_sizing vs 部位 sizing — and in every one the persisted rule text is already byte-identical to the prescription, because the persona committed to candidate_0 on the same dimension. The agree arm is genuinely exercised and genuinely byte-stable.
• 4 (day_trader, swing × 2 locales) carry a commitment with no rule_dim to compare against, so they keep the generic line they already had.
• 0 reach the divergence arm. No mock persona diverges — which is exactly why the corpus never exposed this defect, and why the new test brings its own diverging fixture rather than relying on the sweep.

Tests

python3 tests/run_all.py
PASS: all 49 suites passed.

Run on this tree with pandas 3.0.3, yfinance and rich present, so nothing skipped its way to green.

Deliberately out of scope

• No broader rule-memory redesign, per the ruling's closing line.
• No dim backfill for existing condition-slot or dim-less custom commitments, and no change to _slot_commitment's deliberate drop — that would be the new memory object step 4 forbids. Those users get the fail-safe.
• #546's rule_pending path is unchanged for a first review; nothing about the first-review experience moved.
• The agree arm quotes the commitment verbatim and therefore does not translate it. That matches the existing product posture (_reconciliation_lines already prints the persisted rule as written, and a user's own words are never translated); a returning user who switched locales mid-history is the same pre-existing case that opener already has, not something this cut introduces.

🤖 Generated with Claude Code