docs(pm-dispatch): extract incident case law into references/ — pure byte-identical move (#5925 item 7, #7341 PR B) - #7370

Fixes #5925
Part of #7341

PR B of the three-PR #7341 batch. Stacked on PR A's branch (#7357) per the batch card's load-bearing ordering — the extraction's cut points are re-derived from the post-A file, so this diff shows A+B until A merges; the B-only commits are c187173 (the extraction) and merge 724b810 (propagating A's item-10 follow-up).

What moved (pure move, byte-identical)

9,941 bytes of narrative moved verbatim into two new on-demand reference files; at every cut point the main file keeps the rule + a one-line anchor + a pointer naming the exact section:

references/incidents.md opens with the append-only note: new case law lands there by default, existing entries are never rewritten, section titles are the pointer anchor texts — the main file's growth rate for narrative goes to zero (the #5925 ruling's stated intent).

Numbers

• Bytes moved verbatim: 9,941 (script-verified, see below)
• Main file at the extraction commit: 3,098 → 2,997 lines (235,023 → 226,325 bytes); net −8,698 bytes after ~1.2 KB of new anchor/pointer text
• After PR A's item-10 follow-up merged through (724b810), the same pair reads 3,112 → 3,011 lines (236,095 → 227,397 bytes) — the extraction itself is unchanged; item 10 touches sections no chunk came from
• New files: incidents.md 140 lines, compile-surfaces.md 27 lines
• Bytes remaining in main (current branch head): 227,397

Purity discipline and its verification

Chunks were script-extracted, not retyped: an extraction script sliced each block out of the file by unique start/end anchors and wrote it to disk; the assembly script pasted those exact bytes into the reference files. A verification script then asserted, for every chunk: (1) contiguous byte-identical presence in the pre-move blob (PR A's HEAD), (2) contiguous presence in exactly one reference file, (3) absence from the post-move main file. 8/8 OK.

Kept-side adjustments (declared — these are the anchor/pointer lines, outside the moved bytes): each cut point's lead-in sentence had its dangling colon or "下表/见下" deixis re-pointed at the reference section; no moved byte was reworded. Two blocks were cut at sentence boundaries inside a paragraph (blocks 3 and 5), so the kept halves rejoin with a pointer sentence.

Deliberately NOT moved, per the "IF and only if the move can stay pure" condition: the Operational-notes bodies at large (rule and evidence interleave at sentence level — a pure move cannot separate them), and the step-5 two-blindness explainer (#6371/#6218), whose own governing sentence declares it the PM's step-7 review-question basis; moving it would have required rewriting that kept sentence's meaning, failing the pure-move test.

The line-count ratchet goes in PR C, not here: this PR is a pure move by discipline, and the ratchet is tooling + workflow wiring (C's class), which also lets the ceiling be set against the landed post-B count.

Maintainer authorization (verbatim, untranslated)

关于项目经理的技能,接受你的优化意见,并和前面两个优化任务集中处理。派发使用云卡片,用 Fable 5 模型

#5925 item 7's own approval (maintainer, 2026-08-06, recorded on #5925): 主文件留「规则 + 一行案例锚点」,只把长叙述案例抽到 references/incidents.md;纯搬移零改写;排在全部文本增补之后。The "单独成 PR" half was superseded by the 2026-08-10 batch directive above (recorded on #5925's claim-supersession comment); ordering-after-text-additions is honored via the A→B stack.

Verification

• Byte-identity script: 8/8 chunks OK (pre=true, ref=true, removed=true), 9,941 bytes
• node scripts/check-nul-bytes.mjs — OK
• pnpm check:skill-frame-sync — 4 copies isomorphic (the escalation frame stayed in main, untouched)
• pnpm check:skill-compatibility — 11 files reconciled
• pnpm check:agent-model-declared — os-dev.md → opus
• Dangling-deixis grep for 「下表 / 那张表」: both remaining hits point at tables still in main

.claude/-only ⇒ no changeset; skip-changeset applied.

Generated by Claude Code