docs: record PR #62-#77 and the two open follow-ups (#78, #79)

[StackOverFlow11]

CLAUDE.md's 「当前进度」 stopped at PR #61. Fourteen merged PRs were unrecorded, including the whole factor-evaluation contract layer (analytics/eval/) and eleven reproduced minute factors. This is the file every session loads as ground truth, so the gap would have cost re-derivation or, worse, action on superseded conclusions.

Only CLAUDE.md changed (+90 / −4 net, two commits). No code, config, test or other doc touched.

Per-item diff — entries ADDED

#	entry	source of every number
1	因子评估契约层 analytics/eval/ + 十一个分钟因子复现（#63–#73） — three-axis verdict (Predictive × Incremental × Tradable), asymmetric gate, unknown never convicts, N_eff CI lower-bound gating, exploratory cap at Watch; 11-row table (sign / IC / ICIR / turnover / net 1× / no_book·with_book verdict); three loop-level conclusions; the two methodology rules	RESULTS_post_pr75…md v0.9 reference table (verdicts, IC, ICIR, turnover) + factors/ledger.md (net 1×, families, per-factor conclusions) + gh pr view 63
2	契约 v0.8 + v0.9（#74） — defect 1 (aligned spread added cost back; worked case gross 0.000125 / net −0.001983 → cost 0.002108 → aligned −0.002233, and two of four sites feed the Tradable gate); defect 2 (magnitude-sensitive statistic gating a rank axis → per-date Spearman); v0.9 three-valued N_eff CI gate + "UNKNOWN never rescues"; 3 verdict cells changed; post-merge 11-factor reconciliation	HANDOFF_2026-07-21.md §1 + gh pr view 74 + RESULTS…md
3	⚠️ the reconciliation script that could not fail — compare_postmerge.py read the file _baseline_table.py had just overwritten, so old == new held by construction; reproduced in a sandbox with ic=1.0 vs ic=999.0; authoritative baseline is scratchpad/baseline_reports/, not baseline_rows.json	RESULTS…md §"A check of mine that could not fail"
4	执行价 VWAP + 分钟收益复权（#75） — VWAP fills, gate compares the executed price (LOCKED vs OPENED), 1e-6 tolerance with its optimistic, bounded residual, (raw·af)/(raw·af)−1 holding returns, missing adj_factor never defaulted to 1.0, incidence 8.41% / 9.75%	HANDOFF…md §2 + gh pr view 75
5	报告披露修正（#76） — the report misstated the check it performed; three counters computed and thrown away; why the old I5b fixtures were structurally blind (VWAP ≡ close)	gh pr view 76
6	adj_factor 下降质量检查（#77） — threshold 1% with the full-cache sweep table (7898 events / 4479 symbols), why not 0.5% (603081.SH undecidable) and not 2% (misses a real event), and how this WARNING differs semantically from check_extreme_returns	commit 8a3eea5 + data/quality/market.py:142-198 docstring (merged code supersedes #77's body, which still says 2%)
7	🔄 闸门描述改单一来源（#78, OPEN） — #76's regex guard claimed no further copy could ship; review ran seven plausible rewordings and all seven escape. limit_basis_phrase() is now the one authored sentence, composed by the feasibility section, the group report and the Limitations bullet; rendered text byte-identical; the regex is kept with its scope corrected to catching a literal revert. Carries the generalizing line: a regex cannot assert "no other sentence makes this claim"; having no other sentence can	gh pr view 78
8	日频除权正确性审计 — seam hypothesis rejected (960 boundary pairs, 12 changes, all positive, all with matching price jumps); −57.27% is the BSE placeholder 1.0000 on one symbol from a single fetch with no price rows; in-universe exposure 1 pair in 603,258, CSI300 zero; contaminated-vs-clean mean (+4.38% vs +2.28%, median right all along)	FINDING_adj_factor_seam_2026-07-21.md
9	I5d / I5e / I5f 在修正引擎上重跑 — four-way attribution table; I5d Spearman +1.0000 → +0.9000, spread halved, Q4 overtakes Q5, attributed by measurement to the adjustment alone; I5e unchanged and more negative; I5f capacity at the real 1M notional: below-1.0x 20 → 0	RESULTS…md §I5f / §I5d / §I5e
10	🔄 十一因子改基准 exec-to-exec（#79, OPEN） — why the rebasis (the only modellable fill is the 14:51 bar; close-to-close credits a closing auction closing_call_proxy cannot simulate without stk_auction_*); 20 of 22 cells unchanged, two moved toward less confidence; the single mechanism (point estimates strengthen, N_eff falls in 9/11); the two coverage denominators kept apart; the liquidity-correlated, non-random exclusion; correctness acceptance (per-date corr median 0.9890 against a 0.90 floor that raises, 5 rows hand-recomputed to exact 0.0)	RESULTS…md + gh pr view 79

Per-item diff — existing lines CHANGED

line	change	why
「开发约定 · Git」 PR list	appended #62, #63, #64–#73, #74, #75, #76, #77 as merged; explicit OPEN clause for #38 (unchanged standing constraint), #78 and #79	gh pr list --state all
质量门	793 passed → 1740 passed, measured in this worktree; new group counts from --collect-only (contract 337 + standard 113 + figures 7 / eleven factors 421 / exec-vwap 46 / market quality 15→21); 20 配置 → 31/31; phase0 re-run. Old per-phase breakdown kept, labelled historical	measured here: pytest -p no:warnings → 1740 passed, ruff clean, 31/31 configs, run-phase0 → ic 0.9600 / annual 0.8408. PR #79 independently measured main at 1740
剩余（已显式披露）	added: eleven factors all EXPLORATORY and none reached Adopt, only two positive at 1× fees; MMP evidence updated; I5f rewritten to the real notional; ex-dividend audit done; #78 and #79 both open; schema guard still not wired to live data-update; HTML compendium still on v0.8	entries 1, 6–10
路线图 · MMP	rewritten: 暂缓 now rests on more evidence — I5d degraded (+1.0000 → +0.9000, +17.15% → +8.65%) and I5e strengthened (−26.77% → −31.82%)	RESULTS…md §I5d / §I5e / §MMP combined read
路线图 · I5f	"10M 名义下约半数容量不足" marked superseded; at the real 1M the tightest trade clears at 1.22×	RESULTS…md §I5f capacity
路线图 · 因子层收尾	new clause: #79 and #78 await merge; revise the HTML compendium; prefer turnover/cost-sensitive portfolio-level validation over more factors	entries 1, 7, 10
方法论 rule ② (entry 1)	sharpened after #78: not "assert no (N+1)th copy exists" — that is what the regex failed at — but author the claim once and have every other site compose it	gh pr view 78

Writing discipline held

• No NAV increase is presented as improved performance. Every re-run entry states up front that the whole move is the ex-dividend bug fix and that the VWAP fill basis is itself slightly unfavourable (−0.18pp).
• EXPLORATORY / 非业绩声明 caveats carried over verbatim; negative results and our own mistakes recorded (entry 3), in the tradition of the existing P2-3 「曾误以为是粒度」 and P3-4 「过拟合签名」 entries.
• Numbers with no traceable source were omitted rather than reconstructed. No PR number was guessed — refactor(qt): author the price-limit gate description once #78 and feat(analytics): score the eleven minute factors on the 14:51 VWAP exec-to-exec return #79 were read from gh pr list after both opened.

Gates

pytest 1740 passed · ruff clean · validate-config 31/31 · run-phase0 ic_mean=0.9600, annual_return=0.8408 · secret scan on both commits' diffs: token value 0, .config.json 0, long hex strings 0 · no attribution lines in either commit message.