fix(factors): jump_amount_corr_20 had no 14:50 cutoff — a real look-ahead (correctness fix)

[StackOverFlow11]

fix: jump_amount_corr_20 had no 14:50 cutoff — a real look-ahead (correctness fix, standalone)

This is a correctness fix, deliberately NOT part of the factor-layer refactor (design amendment A3): it changes a published factor's values, and the authorization comes from the research side (the owner's standing 2026-07-19 mandate: "全日数据因子一律截断到 14:50……授权直接截断，不必逐因子再问" — this factor predates that mandate and is the only one that missed it).

The defect

jump_amount_corr_20 was the only one of the eleven minute factors with zero references to decision_time / prepare_visible_minute_bars (the others: 4–10). The old runner fed it the whole day's bars (end + 23:59:59). Under the exec_to_exec return basis, day d's entry anchor is the 14:51 bar's VWAP, while the factor's day-d value used d's 14:50–15:00 bars ⇒ its information set crossed its own entry anchor. Reproduced independently by two parties with different symbols/windows: perturbing post-cutoff bars moved 1,373/1,373 and 1,477/1,477 cells for this factor and 0 cells for all other eleven; the sharper cut (perturbing only bar_end >= 14:51) moved 1,363/1,363 — the signal saw the very minute it would have traded in, and the nine minutes after. Its published artifacts state spec.decision_cutoff="14:50:00" / return_basis="exec_to_exec" with verdict Watch, predictive axis PASS — the report described a check it did not perform.

What this PR does

• Adds the 14:50 cutoff to the factor (same prepare_visible_minute_bars machinery as the other ten), plus a generic intraday-boundary guard covering all 11 minute factors + mmp_ew (coverage closure by directory walk, not a hand-written list).
• Carries the correction as structured data (JSON top-level corrections, exactly one entry; CORRECTION_FIELD_MAX_CHARS=2000, over-long raises rather than crops), not as capped prose. Eval contract v1.0 → v1.1 (to_version must equal the spec's own version ⇒ spec.version alone discriminates which side of the correction an artifact is on).
• Lifts the intentional loud block on this factor's exec evaluation (deny-list entry removed; the cross-PR interlock test test_decision_cutoff_visibility.py failed verbatim with the two required actions before that was done).
• Layout fix (review HIGH): figures.py's definition band now sizes by the branch the dashboard actually renders — meta_y = 0.20 if block is not None else 0.10 means daily capacity 5 lines / intraday capacity 4 lines (measured on the real GridSpec geometry), but the old single DEFINITION_MAX_LINES=5 was measured on the daily branch while the exec path hands the renderer an is_intraday=True spec. Result was 9/11 factors' exec dashboards overlapping. Now DEFINITION_MAX_LINES_DAILY=5 / _INTRADAY=4, and test_definition_band_layout.py is parametrized over both forms — the guard previously only fed the close form, i.e. it protected exactly the half that was not broken. Mutation evidence: INTRADAY back to 5 → 11 tests red; revert → green.
• Census guard fix (review MEDIUM): test_the_correction_carrier_is_never_truncated_on_disk was vacuous — MUST_NEVER_TRUNCATE={"corrections"} never intersected the leaf paths _leaves() emits (corrections.date, …). Now prefix-matched on corrections., with mutation evidence: truncating corrections[0].superseded turns it red and names the right field; the old exact-set logic stays green on the same tampered corpus.

Restated results (run, not inferred)

Verdicts on both bases are Watch/Watch, unchanged (exec: IC −0.030840 → −0.030539, N_eff 1162 → 1178). Discipline: the unchanged verdict is a measured outcome of the full re-run, not inferred from the pre/post-cutoff pearson r=0.9993. And the old artifacts are not retroactively validated: every emitted cell was contaminated; the aggregate rank-IC moved little only because the polluted bars are ~4.5% of the 20-day pooled window and rank IC caps per day — that is dilution, not harmlessness.

Two review findings the review itself retracted (recorded here so the record is not distorted)

1. verdict.reasons has no length cap. It goes through sanitize_text, not sanitize_payload; on-disk artifacts contain an untruncated 1,263-char note and 401-char reasons[]. The earlier "16 characters from the 200-char cap" alarm was a false alarm produced by generalizing from a single sample. What stands: six payload fields ARE capped (census hits 44/44 · 44/44 · 44/44 · 44/44 · 22/44 · 20/44); the only genuinely critical one is coverage_bias_bad_vwap (158/200, headroom 42).
2. The PNG overlap is both a pre-existing defect and a this-PR regression — about different factors. analytics/eval/figures.py had exactly 1 commit in git ⇒ the renderer was always like this (frozen eval_valley_price_quantile_exec_no_book_dashboard.png shows three layers of text stacked). But jump's description growing 323 → 702 chars (3 → 5 lines) was a regression this PR once introduced, then pulled back out via the structured carrier (description now 352 chars). Correct phrasing: the renderer always had the defect; this PR briefly dragged jump into it and then dragged it back out.

Expected downstream signals (pre-registered, not regressions)

• hand_anchors_d2.json will report all_ok_frozen14: False (70 rows, 5 mismatches, all and only jump: hand-computed is now cut off vs frozen panels_d2 which is not). This is the correct signal; the frozen panels are kept as-is, and artifacts/refactor_baseline/pr_c_cutoff_fix/ holds the "old engine + cut input" reference for the refactor-only comparison.
• The 9/11 exec-dashboard overlap for the other factors is a pre-existing renderer defect; their PNGs re-render at the C5 full rerun.

Review & gates

• Branch head 497e50d: cutoff + structured carrier + contract v1.1 (previously reviewed) + merge of main (D4c era, clean) + the two blocker fixes + NIT wording. The post-merge commits were adversarially reviewed by a separate executor: merge face (both sides' semantics in binding.py preserved), independent geometry re-measurement (capacity 5/4 confirmed to the pixel), both mutation evidences re-run, NIT numbers independently verified, no smuggled changes — APPROVE.
• Gates rerun independently by the lead: pytest 2389 passed / 5 skipped (4 census corpus tests skip in worktrees with no artifacts corpus — verified 4/4 PASS against the real corpus; 1 gitignored-baseline-bytes skip), ruff clean, run-phase0 demo anchor ic 0.9600 / annual 0.8408, validate-config 31/31, secret scan (three-dot diff) 0, exec_baseline_freeze --verify 77/77 @ 45c14aa.
• Jump's 4 artifact cells regenerated from the branch via a git archive export (cache-only, stk_mins_live_calls=0, zero cache writes): all 9 JSON/MD byte-identical, PNGs unchanged — correct and expected, since jump's 3-line description fits under both the old and new caps (it was one of the two factors never overlapping). Exec dashboards visually inspected: definition band / metadata / minute-block rows clean.