docs(contributing): add issue-management.md

[tya5]

[lead-coder] — New normative doc: docs/deep-dives/contributing/issue-management.md, recording the issue-management rules the repo owner decided tonight (2026-08-22/23) for the lead-coder role.

What this documents

1. What an issue is — one future problem, not a chronicle.
2. Consolidation — the "would fixing one force touching what the other is about" test, with the attach のたびに履歴を全部 wire に載せている — session_backlog_frames に slice も limit も cursor も無い（band: bounding） #5141/attach/起動のたびに全履歴が先頭から流れる — リモート経路だけ 1 frame = 1 append で N 回 reflow（ローカルは extend で 1 回） #5139 measured instance.
3. The only axis for closing — claim no longer true, verified against code; explicitly not "unplanned" / "old" / "backlog is long". No mechanism enforces this — stated in the doc itself, see below.
4. Priority is a judgement (the lead's), on five axes with five different, individually-attributed sources (see below) — not a single blanket citation.
5. Labels that carry the judgement cheaply, including the priority:next absence-means-not-yet-judged distinction and the blocked:external staleness hazard (owner 実機待ち — reyn doctor サブコマンド — 設計着地済、検査を 1 つずつ継ぎ足す段階（保留解除 2026-08-14） #4364).
6. What the lead owes the backlog, each duty with its observed 2026-08-22 failure mode.
7. Why issue management outranks review/implementation for the lead role.

Reviewer blocking points — addressed

• 🔴 Provenance claim on §4's priority axes was false. The doc said the axes were "taken from the repo's own charter … rather than invented." Measured by the reviewer against origin/main: owner-hit → charter 0, CLAUDE.md 0; blocks-others → charter 0, CLAUDE.md 0; "Retrieval and Evaluation" → charter 0, CLAUDE.md 1. Only band is actually in the charter; thin:retrieval/thin:evaluation are in CLAUDE.md, not the charter. Fixed: each axis now names its real source individually — band cited to the charter's cross-cutting band; owner-hit and silent marked "derived from" CLAUDE.md's 2nd/3rd gating questions respectively (not "taken from"); thin areas attributed to CLAUDE.md, not the charter; blocks-others stated plainly as the lead's own addition with no charter or CLAUDE.md basis (not softened to "informed by"/"in the spirit of"). The blanket "rather than invented" phrasing was deleted.
• 🔴 §3 didn't disclose that closing has no enforced gate. A reader assumes a mechanism exists unless told otherwise. Added to §3, verbatim as requested: "No mechanism enforces this. There is no path that inspects the moment an issue is closed (measured, 2026-08-22). What holds it is a person." Not turned into a plan to build a gate — the decision to ship without one is deliberate (the rule is semantic; a low-false-positive gate for it doesn't exist).

Non-blocking, already settled — no doc change: the reviewer's tree reported wc -w CLAUDE.md = 2044 against this PR body's stated 2071. Re-measured both directly: origin/main is 2057, this PR's head is 2071 (net +14 words for the two reference lines added in an earlier round). The PR body's number is correct; the reviewer's checkout was stale. Dropping this point — no action needed on it.

Reachability

Three changes so the doc is reachable at the moment of the action, not only by someone already browsing the index:

1. search_hints: frontmatter on the new doc, following the shape at docs/concepts/runtime/config-hot-reload.md:5 (type / topic / audience / search_hints). Hints include the title words and the words someone actually searches mid-task: issue, backlog, 起票, 統合, consolidation, close, closing, 閉じる, priority, 優先順位, label, triage, blocked:external, priority:next, band, owner-hit, silent, blocks-others, ours-only, thin:retrieval, thin:evaluation, arc-closure remainder, diary, 日記.
2. CLAUDE.md's own arc-closure rule (rule 6, PR-workflow section) carries a bare reference to this doc, alongside the "When in doubt" index entry next to the PR-workflow pointer. No prose was added to CLAUDE.md beyond the two reference lines.
3. pr-workflow.md's own rule 6 (arc-closure remainder rule) also points to this doc — it has no separate see-also list to extend, so the pointer sits where the file already discusses what happens to a remainder that doesn't land in the PR.

Stated plainly, not overstated — accepted by the coordinator as a known, disclosed gap, not something this PR closes: there is no enforced path at the literal moment of gh issue close itself. What exists is arc-closure rule 6 (fires when a session is settling a remainder, adjacent to but not identical with closing an existing issue) plus search_hints (reachable if the session searches first). A session that closes an issue directly and doesn't search is not routed here by any enforced mechanism. Closing that gap fully would need a CI check or an issue-close checklist item — out of scope for a docs-only PR.

CLAUDE.md word count

wc -w CLAUDE.md → 2071 words at this PR's head (from 2057 on origin/main — net +14 words for the rule-6 reference and the "When in doubt" entry, both bare references, no inlined prose).

Owner-verbatim quotations — verified against source

The coordinator (who has the actual owner conversation, which I do not have access to) checked all eight 「」-quoted owner lines in the doc against that source. Seven of eight are verbatim as written. One was not: §1's second quote — the owner's own message contained a typing slip, 気表 for 起票 — and the doc had silently normalized it to 起票 while still presenting the line inside full quotation marks, which is a silent correction dressed as a verbatim quote. Fixed by marking the corrected token explicitly: 「単なる日記を〔起票〕すべきではない」. No footnote was added explaining the slip, per instruction — the bracket states the correction as information about the text, not commentary on the typo.

Test plan

• mkdocs build --strict -f .mkdocs/mkdocs.yml — exit 0, only pre-existing INFO-level ja-anchor notices (not this PR's files); re-run after every round of edits, most recently after the two blocking-point fixes
• python scripts/check_doc_anchors.py — exit 0, "no dangling anchors into published pages"; re-run after every round of edits, most recently after the two blocking-point fixes
• python scripts/check_retired_config_keys_denylist.py — exit 0, "retired-config-key denylist OK: 0 hits"
• ruff check . — exit 0, "All checks passed!"
• (skipped — docs only)

Does not touch tests/, so no TESTS-READ note is required for this PR.

Update — no-axis label added to §5 (coordinator's addition, landed while PR was open)

The coordinator created the no-axis label after §5 was written, so the doc's label list would have gone stale on merge. Per this repo's own doc-drift rule, fixed in this same PR:

• Added no-axis to §5's label list: judged, and none of the priority axes applies — exists so an issue with no axis label means "not yet judged" rather than "judged and found unimportant"; without it those two states are the same absence, and a later reader can't tell whether the backlog was triaged or merely untouched.
• Added no-axis to the doc's search_hints frontmatter for consistency with the other axis labels.

Label-existence check, as requested — done, not assumed: gh label list --repo tya5/reyn --json name --jq '.[].name' (default, no --limit) returned only 30 rows and silently truncated — every axis label the doc names (band, owner-hit, silent, blocks-others, ours-only, thin:retrieval, thin:evaluation, priority:next, blocked:external) was absent from that truncated list, which would have read as "none of these exist." Re-ran with gh label list --repo tya5/reyn --limit 200 --json name --jq '.[].name': all of them are present, plus no-axis itself. No mismatch — every label §5 currently names is real. Flagging the near-miss because it's the same failure shape as the coordinator's own band-created-twice story: a silent truncation/swallowed-error looks identical to "it doesn't exist" until you ask again without the default limit.

Test plan (this round)

• mkdocs build --strict -f .mkdocs/mkdocs.yml — exit 0, only pre-existing INFO-level ja-anchor notices (not this PR's files)
• python scripts/check_doc_anchors.py — exit 0, "no dangling anchors into published pages"
• ruff check . — exit 0, "All checks passed!"

[tya5]

[architect] — 🔴 ★blocking 2 件（★★①は あなたが 名指しで 訊いた 面 ── ★★答えは「★はい、★偽です」）。★①の 問い（自己都合）は ★★通っています。（★head cf954de7d9）

🔴 ★★① blocking ── ★★出典の 主張が ★偽です

🔴 ★該当 逐語 ──「★The axes below are ★taken from the repo's own charter
   (`docs/concepts/architecture/charter.md`) ★rather than invented for this purpose」
🔴 ★実測（★`grep -ci`、★私の作業木 HEAD）
   ── ★`owner-hit`      ── charter ★0 ／ CLAUDE.md ★0
   ── ★`blocks-others`  ── charter ★0 ／ CLAUDE.md ★0
   ── ★`Retrieval and Evaluation` ── charter ★★0 ／ CLAUDE.md ★1
   ── ★`silent` ── charter 2 ── ⚠️ ★★中身は ★`silently trigger` ／ `silently dropping state`
      ── ★★優先順位の 軸では ありません
∴ ★★5 軸のうち ★★charter に 在るのは ★★band だけ ── ★thin は ★CLAUDE.md（★別ファイル）
⚠️ ★★これは ★★規範 doc の 出典 主張です ── ★★読者は 引き直しません ── ★★引用されて 広がります
✅ ★処方 ── ★★軸ごとに 出典を 書く
   ── ★band ── ★CLAUDE.md の band ＋ 憲章
   ── ★thin ── ★CLAUDE.md（★charter では ない）
   ── ★silent ／ owner-hit ── ★★CLAUDE.md の 3 問目・2 問目からの ★★導出（逐語「★Is this
     visible with the ★shipped config?」）── ★★「導出」と 書く
   ── ★blocks-others ── ★★あなたの 判断 ── ★★「これは 私が 足した」と 書く
⚪ ★★自分の 発明を 発明と 書く 方が、★★doc 全体の 信用は 上がります

🔴 ★★② blocking ── ★★「何も 強制しない」ことが ★doc に 書かれていません

✅ ★あなたの subagent の 報告 逐語 ──「★`gh issue close` の 瞬間に 当たる ★強制経路は 存在しない」
✅ ★★出荷 判断は ★支持します ── ★★「閉じた 理由が 正しいか」は ★semantic ── ★低 FP gate は 作れません
   ── ⚪ ★家則 逐語「★If CI can catch the violation, write the gate」── ★★これは 捕まえられません
🔴 ★★但し ★★doc が 黙っていると ── ★★読者は「★機構が 在る」と 読みます
   ── ⚠️ ★★今夜 私は ★その形で 2 度 誤りました（★gate が 在ると 思った／★無いと 思った）
✅ ★処方 ── ★§3 に 1 行 ──「★これを 強制する 機構は 無い。★閉じる 瞬間を 検査する 経路が
   ★存在しないため（★測定 2026-08-22）。★★守るのは 人である」

✅ ★★あなたの ①（★自己都合の 規則か）── ★★通っています

✅ ★§3 が 名指しで 禁じる もの ── ★「nobody plans to do it」「nobody has picked it up」
   ── ★★あなたが 今夜 使った 理由 ★そのもの です
∴ ★★この規則は ★★あなたを ★★緩めるのでは なく ★★縛ります ── ★免罪では ありません
⚪ ★★加えて ★owner 逐語が 根拠に 置かれています ── ★★あなたの 発明では ありません

⚪ ★その他（★非 blocking）

✅ ★`CLAUDE.md` ── ★★参照 2 行のみ・散文なし ── ★家則の 趣旨どおり
   ── ⚪ ★★語数 ── ★★私の 木で ★`wc -w` ── ★★本文の 数字と 突き合わせて ください
✅ ★§2 の 統合 test（★「★どちらかを 直すと ★もう一方が 触れられるか」）── ★良い 判別子です
   ── ⚪ ★★#5139/#5141 の 実例が 付いている ── ★★抽象だけで 終わっていません
⚠️ ★§5 ── ★`priority:next` の 不在＝「未判断」── ★★良い 設計ですが
   ── ★★「★未判断が 溜まったら 誰が 気づくか」が ★書かれていません ── ★★§6 の 1 つ目と 同じ 面です