[META] Codify append-vs-modify discipline as plugin-level principle

[kiki830621]

Problem

Original text (from /spectra-discuss session 2026-05-25,user 切到 append-vs-modify 議題後 explicit pivot):
「我最近開始思考的是,不斷 append 會不會不比直接修改內容好」
— Source: spectra-discuss session conversation (本對話即時 capture)

IDD 內部其實已經是 hybrid append + modify(不是純 append),但 hybrid 規則 implicit、case-by-case 演化、從未被 codify。每次遇到「AI 想表達新狀態但不能直接編老 AI-authored 內容」就加 supersession workaround,workaround 數量正在增加。本 issue 是強迫 explicit 決定該不該把 implicit 規則升上 first-class principle。

Type

refactor — META / plugin design principle codification(類比 #144 IC_R011 「AI design over-abstract」 的 plugin-level principle file pattern)

Expected

• Explicit Append vs Modify discipline 文檔化為 plugin-level rule(可能落地形式:plugins/issue-driven-dev/rules/append-vs-modify.md 或併入既有 manifesto)
• 提供 mapping table:artifact type → modify / append / hybrid 行為對照
• 解決痛點:Strategy/Plan checkbox supersession 改走 gate-logic「only-look-at-latest」(Path C)還是允許 AI 直接 PATCH 老 AI output(Path B);決定 IC_R007 與 「AI self-edit」 的 boundary

Actual

• Discipline 是 implicit,scatter across 7+ sites(comments / issue body / CHANGELOG / canonical ref / SKILL.md / tasks.md / supersession layer)
• 每個新 IDD feature 都在加 ad-hoc supersession workaround(#515 / [design] IC_R011 follow-up checkpoints should default to filing, not asking — past suggestions are disappearing into conversation #148 / [docs] idd-verify Step 5b lacks canonical "Rule (SHALL/SHOULD)" framing — IC_R011 spec consistency gap #149 retroactive summary / IC_R011 audit trail accumulation),沒人問 root rule
• 新人 / 3 個月後讀者要 archaeology 多個 comment 才能 reconstruct current state

現狀證據(7+ sites,本 session 親自觀察)

Artifact	行為	證據
Comments	純 append + 內部 PATCH	diagnose 後續 PATCH audit blocks(Layer V + Sister + Distribution Sync)
Issue body 上半(Problem/Type/Expected/Actual)	Immutable(IC_R007 verbatim preserve)	/idd-issue Step 3 contract
Issue body 下半(Current Status)	Modify(REPLACE below ---)	/idd-update Step 5
CHANGELOG.md	純 append	[2.72.0] 加在 [2.70.0] 上方
Canonical ref .md	Modify(in-place rewrite)	ic-r011-checkpoint.md 301→397 是 Write-replace(本 session 親手做)
SKILL.md	Modify(refactor in-place)	6 SKILL.md 從 inline ~50 行 → cite ~15 行(本 session 親手做)
tasks.md	Modify(checkbox flip)	[ ] → [x] via spectra task done
Strategy/Plan checkboxes in 老 comments	Append-only with supersession hack	#515 fix:post 新 ## Implementation Complete override 老 Strategy/Plan
Closing summary 內 cross-link	Modify(Step 3.5 inline replace BEFORE publish)	/idd-close Step 3.5
Spec archive	Hybrid(delta promote to main + snapshot keep)	spectra archive 預設 sync delta → main

隱性原則(從上表反推)

human-authored content    → never modify (IC_R007 verbatim preserve)
AI-authored decision      → append (decision history matters)
machine-readable state    → modify (only current state matters)
AI-authored stale state   → supersession hack (post newer to override older) ← 痛點所在

痛點具體 evidence(本 session)

1. #515 fix:idd-close Step 0 gate 一直誤判 Strategy/Plan 老 checkboxes 為「未完成」→ 加 supersession bridge(Implementation Complete > Checklist 全 [x] 視為 canonical state of truth,跳過 Strategy/Plan scan)。
2. [design] IC_R011 follow-up checkpoints should default to filing, not asking — past suggestions are disappearing into conversation #148 closing remediation:本 session 必須手動 post 一個 synthetic ## Implementation Complete comment 才能觸發 #515 supersession → gate 通過 → close。
3. [docs] idd-verify Step 5b lacks canonical "Rule (SHALL/SHOULD)" framing — IC_R011 spec consistency gap #149 retroactive closing summary:被 commit-body trap auto-close,事後 post 「(retroactive — auto-closed via commit-body trap)」標記的 Closing Summary。Audit trail 同時有早期 auto-close 跟事後 summary,讀者要自己分辨哪個是 final state。
4. IC_R011 audit trail accumulation:### Sister Concerns Filed / ### Closing Follow-ups Filed / ### Distribution Sync / ### Residue Acknowledgement 都 append 到同一 comment,而非 REPLACE 一個 "Final Audit" structured section。Comment 不斷膨脹;3 個月後讀者要分辨「這 4 個 block 是真的 final 還是還會繼續長」。
5. Canonical ref 文件 growth:ic-r011-checkpoint.md 在本次 spectra change 從 301 → 397 行 append 上去(Section 7、8 新增)。這個是 modify-friendly artifact,所以沒 friction,但證明 IDD 在 docs / SKILL.md 已經是 free-modify;痛點集中在 GitHub-side comments / issue body。

3 個 candidate path(spectra-discuss 結論)

Path	機制	代價
A: Status quo(append + supersession)	老 comment 留著 + 新 comment override	讀者 archaeology;3 個月後翻舊 issue 像考古;history 完整但「current state 在哪」不明確
B: AI may modify AI's prior output	idd-implement 結束時直接 PATCH 老 ## Implementation Plan checkbox 為 [x],或 PATCH 老 Strategy 加 ~~strikethrough~~ + 「(superseded by #NEW comment)」	失去「decision-at-time」history;但 single source of truth;需要明確的 IC_R007 vs AI-self-edit boundary
C: Refactor gate semantic	把 Strategy/Plan checkbox 設計成「初始 snapshot,not gate source」;gate 永遠從新 Implementation Complete 讀(#515 已部分實作)	不改文件 mutability,改 gate logic 預設 only-look-at-latest;最小破壞;但仍留 archaeology cost

spectra-discuss tentative inclination:C 加部分 B(C 解 90% 痛點 + B 用於明確「AI retract self」case)— 留給本 issue 的 diagnose / propose 階段正式 decide。

Sister Concerns

• [META] codify 'AI design 階段過抽象,需要 third-party surfacing' as plugin-level principle (residue from #135) #144(codify「AI design over-abstract」as plugin-level principle)— 同 pattern:implicit AI behavior → user explicit surface → codify as plugin principle。本 issue 是 instance idd-all: direct-commit mode 加 passive warning (dirty tree / on default branch) #2 of this meta-pattern。
• #515(supersession bridge fix)— 本 issue 是 root-cause-level codification,#515 是 surface fix。本 issue 可能 retroactive 改變 #515 的 mechanism choice。
• [META][unattended] Clarity Surface annotation interaction contract with /loop, /idd-all (sister concern from #135) #137(unattended-mode Clarity Surface contract)— 被 deprioritized 切到本 issue(user 在 spectra-discuss 中 explicit pivot):「(2) 先 file 這個 append-vs-modify 議題 → 再回 [META][unattended] Clarity Surface annotation interaction contract with /loop, /idd-all (sister concern from #135) #137」。本 issue ship 後 resume [META][unattended] Clarity Surface annotation interaction contract with /loop, /idd-all (sister concern from #135) #137 discussion。

Open Questions(deferred — 留給 diagnose/discuss 決定)

1. AI 可不可以 modify AI 自己的舊 output?(Path B)若可,要什麼 boundary?(只 PATCH checkbox? 還是可以加 ~~strikethrough~~ + supersession note?)
2. 跟 IC_R007 「human content verbatim preserve」的 boundary 怎麼界定? AI 引用 human content 後 AI render 的詮釋是 AI-authored,可被 AI 自己 modify 嗎?
3. 保留 append-only,改 gate logic「only look at latest」?(Path C)idd-close gate 預設只看最新 ## Implementation Complete > ### Checklist(#515 部分實作)— generalize 到所有 gate(idd-verify / idd-update)?
4. IC_R011 audit trail blocks 是否該收斂成 single Final Audit section + REPLACE,而非各自 append?
5. idd-edit 已經 modifies content — 它跟 append-vs-modify discipline 的關係?是 escape hatch 還是該被 discipline 約束?
6. 直接 file 為 plugin principle(類比 IC_R010 / IC_R011)還是 codify 為 manifesto 一節?

為什麼 file 這個 issue(meta-justification)

第 4 條 supersession workaround 數量已經 ≥ 5(從本 session 親身證據)。每個新 IDD feature 都在加 workaround,而沒人在問「這條 implicit rule 該不該升上 explicit principle」。本 issue 的目的是強迫做這個決定,而不是繼續在每個新 friction 加 ad-hoc bridge。

Residue

NSQL §4.6 residue(由 idd-issue 自我聲明,後續 diagnose 階段補充):
本 issue 涵蓋「IDD 內部 append vs modify discipline」這個 decision space,但不涵蓋:

• IDD 與其他 plugin 的 cross-plugin append/modify policy(本 issue scope 限 IDD 自己)
• 過去已 ship 的 supersession workaround(#515 等)是否該 retrofit(留給 implement 階段 decision,本 issue 只 codify principle)
• 跨 session 的 comment 累積最佳實踐(屬於 GitHub UX 議題,不是 IDD 設計議題)
• 本對話即時討論的 Path A/B/C trade-off 細節(diagnose 階段重新評估)

(本 issue 是 META/principle 議題,不是 single behavior change — 屬 Spectra-tier candidate per IC routing,但最終 routing 由 /idd-diagnose 決定。預期 verdict: Spectra(Layer 2 published principle + Layer 3 修改既有 spec normative behavior 多處)。)

Clarity Surface (dismissed)

Type	Source	Suggested canonical	Status
(dismissed)	Step 4.6 was deferred initially (META-recursion: issue body uses IDD plugin's own canonical terms supersession / IC_R011 / append / modify / audit trail / IC_R007 that user explicitly defined in immediately-prior /spectra-discuss session 2026-05-25). Step 0.5 gate at /idd-diagnose triggered hard-refuse on deferred row; dismissed with canonical §1.4(a) unactionable-observation reason: META-issue about IDD's own discipline, no domain-vocab mismatch exists to clarify. Reviewed body: 0 mismatches found.	n/a	dismissed

Audit dance: deferred at /idd-issue Step 4.6 → REFUSE at /idd-diagnose Step 0.5 → dismissed with §1.4(a) reason → gate passes. This is the canonical "deferred-with-reason → gate-refusal → explicit-dismiss" loop the spec designed for.

Linked-Context Siblings Filed (v2.48.0+ #529, per IC_R011 v2.72.0+ default-flip)

Filed	Title	Source	Disposition
#151	[Bug] Step 0.8 commit-body auto-close trap protection gap	/idd-close cluster #148 #149 session 2026-05-25 (#149 被 closes #149 body trap auto-close)	filed per canonical §1.1 file-by-default — user explicit confirm in Step 4.7 picker

Sweep scope: body draft + recent session conversation (~hours of /spectra-discuss + batch diagnose). #151 is structurally orthogonal to this META principle issue but informationally adjacent (Path B remediation choice depends on #150's principle decision).

---

Current Status

Phase: closed
Last updated: 2026-05-25 by /idd-update (via /idd-close)

Key Decisions

• Action-scoped reframe (user reject actor-based exemption): 7-category taxonomy (state-field-update / bounded-section-replace / audit-block-append / inline-replace-before-publish / verbatim-preserve / append-only / free-rewrite)
• Spec discipline declared in v2.73.0 + v2.74.0 (8 SHALL + 16 scenarios); /idd-edit BREAKING runtime enforcement DEFERRED to [feature] /idd-edit runtime enforcement — proper proposal after 3-iteration bash failure (R1/R2/R3 on #153) #154 per R4 Path (c) decision after 3 verify iterations exposed bash-incremental impl introduces new bugs each pass
• Path C authoritative_source gate-logic generalization across 4 sites (idd-close Step 0 / idd-verify checklist scan / idd-update body sync / idd-implement Step 5 Checklist Sync)
• Reason pattern registry section added to rule (cited by [META][unattended] Clarity Surface annotation interaction contract with /loop, /idd-all (sister concern from #135) #137 work)
• Cluster PR cluster: action-scoped modify discipline + /idd-clarify unattended auto-defer (#150 + #137) #153 (with [META][unattended] Clarity Surface annotation interaction contract with /loop, /idd-all (sister concern from #135) #137)

Scope Changes

• DROPPED from spectra-apply tasks 3.1+3.2+3.3 (idd-edit BREAKING runtime enforcement) → moved to [feature] /idd-edit runtime enforcement — proper proposal after 3-iteration bash failure (R1/R2/R3 on #153) #154

Blocking

• (none — issue closed; [feature] /idd-edit runtime enforcement — proper proposal after 3-iteration bash failure (R1/R2/R3 on #153) #154 inherits runtime enforcement work)

Commits

• b788a73 cluster merge commit (same as [META][unattended] Clarity Surface annotation interaction contract with /loop, /idd-all (sister concern from #135) #137; both issues co-shipped in PR cluster: action-scoped modify discipline + /idd-clarify unattended auto-defer (#150 + #137) #153)

Closing

• Closing Summary: [META] Codify append-vs-modify discipline as plugin-level principle #150 (comment)
• Spectra archive: openspec/changes/archive/2026-05-25-add-action-scoped-modify-discipline/
• Follow-up [feature] /idd-edit runtime enforcement — proper proposal after 3-iteration bash failure (R1/R2/R3 on #153) #154 filed (idd-edit runtime enforcement with R1/R2/R3 evidence + design requirements)

[kiki830621]

Diagnosis

Type

refactor — META plugin design principle codification

Analysis (現狀分析)

為什麼要 codify:IDD discipline 內部已經是 hybrid append + modify,但 hybrid 規則 implicit、case-by-case 演化、從未升上 plugin-level principle。每次新 feature 撞到「AI-authored stale state」就加 supersession workaround。Workaround 數量正增加 + 沒人問 root rule。

現狀問題具體 evidence:

• 7+ sites 各自有不同 mutability behavior(本 issue body 對照表)— mental model 散在各 SKILL.md / canonical ref 各處,無 single source
• ≥ 6 個 supersession workaround instance(#515 / [design] IC_R011 follow-up checkpoints should default to filing, not asking — past suggestions are disappearing into conversation #148 retroactive / [docs] idd-verify Step 5b lacks canonical "Rule (SHALL/SHOULD)" framing — IC_R011 spec consistency gap #149 retroactive / IC_R011 audit accumulation / canonical ref growth / 本 body 被 sed 寫壞復原)— 數量持續增加
• 痛點集中在第 4 條「AI-authored stale state」:Strategy/Plan checkbox supersession 需要 #515 supersession bridge;closing summary retroactive([docs] idd-verify Step 5b lacks canonical "Rule (SHALL/SHOULD)" framing — IC_R011 spec consistency gap #149)需事後補 retroactive marker;IC_R011 audit blocks不斷 append 而不收斂

為什麼是 refactor 而非 feature:不引入新外部 API。改的是 IDD 內部 hybrid discipline 的 codification + 可能 retroactive 修 #515 機制選擇 + 影響後續 IC_R011 / unattended-mode contract / supersession-bridge 設計。

為什麼是 META principle 而非 single-behavior change:本 issue 不是「修一個 supersession workaround」,是「決定 supersession 該怎麼存在」— 影響至少 6 個 site 的設計。類比 IC_R007 / IC_R010 / IC_R011 的 codification pattern,屬同層次 principle。

Impact

影響檔案類別(初估,確切清單待 spectra-propose):

• 新增(1-2 個 plugin rule 檔案):
  • plugins/issue-driven-dev/rules/append-vs-modify.md — principle 主文 + artifact-type mapping table
  • 可能合併到 manifesto.md 一節而不是獨立檔案(待 propose 階段 decide)
• 修改既有 SKILL.md normative behavior:
  • idd-close Step 0 supersession logic — 視 Path 決定 generalize 還是保留現狀
  • idd-implement Step 5 Checklist Sync behavior — 若 Path B 採用,可能 PATCH 老 Strategy/Plan comments;若 Path C,維持現狀
  • idd-verify / idd-update gate logic — 對齊「only-look-at-latest」原則(若 Path C)
• 修改既有 reference docs:
  • references/ic-r011-checkpoint.md audit trail section — 可能收斂多 block 為 single Final Audit + REPLACE
  • 多個 SKILL.md 的 audit trail discipline 段落需引用新 principle
• 使用者流程影響:
  • Sister [Bug] Step 0.8 commit-body auto-close trap protection gap (direct-commit path) #151 Path B remediation 設計直接受本 principle 決定影響
  • Sister [META][unattended] Clarity Surface annotation interaction contract with /loop, /idd-all (sister concern from #135) #137 unattended-mode Option B(unattended_review_pending 新 status)設計受影響(該 design 等於再加 append slot,可能被 Path C 否決)

Strategy

已在 spectra-discuss 累積 3 個 candidate path(本 issue body 內表已列):

• Decision: Path A / B / C / B+C 選擇 — 在 spectra-discuss 收斂 + spectra-propose 鎖定
• 寫 append-vs-modify.md(or manifesto section)principle 主文 + artifact-type mapping table(本 issue body 的 7-site 表 source-of-truth)
• 若選 Path C:generalize idd-close Step 0 「only-look-at-latest」邏輯到 idd-verify / idd-update 所有 gate
• 若選 Path B 部分:codify 「AI may modify AI's prior output」 boundary(只 PATCH checkbox? 或 ~~strikethrough~~ + supersession note?);明確區分 IC_R007 verbatim-preserve vs AI-self-edit
• 收斂 IC_R011 audit blocks 為 single Final Audit section + REPLACE(若 Path C 採用)— 或維持各自 append(若 Path B 採用)
• Update at least 4 既有 SKILL.md 引用新 principle 並對齊行為
• 寫 audit-trail test fixture(scenario:Strategy + Implementation Complete supersession + retroactive close)驗證新 gate logic

Complexity

Spectra(Layer 2 + Layer 3 全命中)

Layer 1 disqualifier check:無命中(不是 narrative / ad-hoc analysis script / typo cleanup;改的是 normative behavior)

Layer 2 necessary condition(命中):

• 改動為 plugin maintainers / downstream IDD plugin consumers / 未來 IDD feature authors 產出 documented contract — 是新 published principle,給 future callers 寫成 spec

Layer 3 confirmation signals(≥ 1 命中,實際 3 命中):

• ✅ 修改既有 published spec normative behavior(idd-close Step 0 supersession / idd-implement Step 5 / references/ic-r011-checkpoint.md audit / 多個 SKILL.md audit trail clause)— 全部是 SHALL clause changes
• ✅ 影響 2+ 既有 specs 需要 consistency-checking — 至少 idd-close / idd-implement / idd-verify / idd-update / ic-r011-checkpoint / sdd-integration manifesto 6 個 spec 連動
• ✅ Architectural decision with long-term maintenance implications — 不是 method-level,是會被所有 future IDD feature author 繼承的結構性 discipline

Risks

Risk	Severity	Mitigation
Path B 引入 modify 操作 → corruption blast radius	HIGH	本 incident(body 被 sed 寫壞)是直接 evidence。若採 Path B,必須要求 atomic modify(non-bash;直接 GitHub API JSON PATCH;有 dry-run mode)+ pre-modify body snapshot to local for recovery
Path C 仍留 archaeology cost	MEDIUM	3 個月後翻舊 issue 仍要讀多 comment 才知 current state。Mitigation:在 issue body Current Status 補 link to latest authoritative comment
Retroactive 改 #515 mechanism choice 破壞 backward compat	MEDIUM	Existing closed issues 已用 supersession bridge 成功 close。新 mechanism 要 backward-compat handle 老 issue。Mitigation:gate logic fork — 看是否有「Implementation Complete > Checklist」就走新邏輯,否則 fall back 老邏輯
idd-edit 已是 modify entry — discipline 衝突	LOW	確認 idd-edit 屬「使用者 explicit modify intent」exception(非 AI auto-modify)。Discipline 約束 AI 自動行為,不約束 user explicit edit
Sister #137 design 必須等本 issue ship	MEDIUM	#137 已在 issue body 標 deprioritized。已有 hard-refuse 暫時 fallback,功能不破
過早 codify(rule-of-three 違反?)	LOW	反 evidence:已 ≥ 6 instances(本 incident 是第 6 個),早 over rule-of-three。User #144 + #150 explicit signal「不再等 instance」。

Residue

NSQL §4.6 residue:

本 issue 涵蓋「IDD 內部 append vs modify discipline 的 codification + 採 Path B/C/B+C decision」。不涵蓋(intentional out-of-scope):

• 跨 plugin 的 append/modify policy — 本 issue 限 IDD 自己;若 che-msg / che-zotero 等 plugin 有類似議題,各自 file own principle issue
• GitHub-level UX 議題 — 例如 comment 累積閱讀體驗、PR review UI、issue timeline navigation;屬 GitHub 產品,non-actionable for plugin
• Spec-level archive policy(spectra archive 的 delta-to-main promote) — 已是 hybrid 但無痛點;若採 Path C 不影響,若採 Path B 也不影響(已 modify behavior)
• 本 incident root cause 的 sed/bash unsafety detail — 屬「shell scripting safety」議題,可以開 sister issue 但不在 META principle 範疇

Audit Notes(meta)

本 diagnose 過程本身觸發 2 個 #150 自身的 instance:

1. Step 0.5 gate 自我 trap:我 file [META] Codify append-vs-modify discipline as plugin-level principle #150 時留下 deferred Clarity row,/idd-diagnose Step 0.5 gate REFUSE。 honest dismiss with §1.4(a) → gate pass。 → 證明 deferred-with-reason 機制 work as designed。
2. Body 被 sed 寫壞:dismiss row 時 bad sed substitution(|| 被當 flag)+ gh issue edit 把整個 body 寫成 1 個 newline。 從 local /tmp/...md reconstruct 復原。 → instance idd-all SLUG derivation: BSD/GNU sed alternation (\| not portable) #6 evidence,加入 issue body live evidence 段。

兩個 incident 都是本 issue 的 recursive validation:本 issue 在 diagnose 自己時就證明了痛點存在。

Vagueness Pre-check

• V1: 2 — body has explicit deliverables (codify discipline as plugin rule file + 7-site mapping table + Path B/C/B+C decision + update N SHALL-tier SKILL.md). Anchor 2 ("Add a button...") vs 3 ("optimize export..."): scope-bounded with concrete deliverable shape.
• V4: 3 — Strategy items mostly conditional ("若 Path C: ...", "若 Path B: ..."), but the Path decision is the purpose of next-step spectra-discuss, not acceptance vagueness. Anchor 3 ("Done when latency feels noticeably better") rather than 4 ("Make the API response cleaner") — decision deferred, not undefined.
• Triggered: no — both axes ≤ 3

Sister Concerns Filed (mid-diagnose, v2.47.0+ #528)

Filed	Title	Source	Disposition
(none new)	—	scanned diagnosis content + session — #151 already filed at /idd-issue Step 4.7 (auto-close trap protection gap); 2 self-trap incidents during this diagnose (Step 0.5 deferred-gate dance + body sed corruption) are recursive evidence of #150 itself (in-scope, written to body §Live evidence + §Audit Notes); shell-scripting-safety candidate is §1.4(a) unactionable (local plugin micro-discipline, not principle)	n/a

Sweep scope: just-posted Diagnosis comment (Strategy + Risks + Audit Notes sections) + ~hours of session log (cluster close #148 #149 + batch diagnose #136-#146 + spectra-discuss #137 → #150 pivot).

[kiki830621]

Verify (via PR #153)

Result: FAIL — 3 BLOCKING bugs + 5 MEDIUM + 5 LOW findings
Full report: #153 (comment)

This issue's findings: see master report — #150 primarily affected by B3 (/idd-edit --replace BREAKING is documentation-only,SKILL.md Step 1/4 lacks enforcement parsing) + L3 (retroactive category labels gap on idd-implement Step 5.7 + idd-verify Phase 4/Step 5b)。

Next: fix loop in same PR per feedback_verify_fix_same_pr。

[kiki830621]

Verify R2 (via PR #153) — DEGRADED engine

Result: PASS-with-followups (all 3 R1 BLOCKING fixed; 4 NEW findings from coordinator self-DA)
Full report: #153 (comment)

Engine note: 5 reviewer Agents rate-limited 2× retries → coordinator self-review fallback (Step 2.5c). Codex still running. Process gap flagged in master.

This issue (#150) affected by NEW findings:

• F2 (LOW) — grep -qF pre-flight + exact-line splice mismatch → silent partial failure on substring-matched section names
• F4 (MEDIUM) — Errata flow regression on user-authored comments (/idd-comment auto-call /idd-edit --prepend-note triggers new verbatim-preserve guard refuse)

User decision needed on fix-in-PR vs follow-up issues.

[kiki830621]

---

⚠ REVISION (Codex audit appended post-master):R2 verdict revised to FAIL — B3 NOT actually fixed。 Documented /idd-edit --replace --scope whole-comment --body "..." happy path is runtime-broken(--scope space form parser sets literal "next";--body space form sets empty)。 R3 fix loop required for B3 parser rewrite。 See master comment update。

[kiki830621]

Verify R3 (via PR #153) — full ensemble

Result: FAIL — R3 fix addressed all 4 R2 findings ✓ but introduced 3 NEW CRITICAL bugs(C1 infinite loop / C2 flag-value-eat / C3 multi-line awk-v BREAK)+ 1 HIGH(--body-file silent empty)
Full report: #153 (comment)

This is 3rd consecutive failure of B3 enforcement implementation。 User decision needed:

• (a) R4 fix loop(more code,risks more bugs)
• (b) Revert R3(R2 was "broken-but-no-loop")
• (c) Revert ALL B3 enforcement(restore pre-[META] Codify append-vs-modify discipline as plugin-level principle #150 state),file as proper follow-up issue,ship R1 awk fix + M5 + L3 only

Master report recommends (c) per lead-minimal + 3-iteration failure pattern。

[kiki830621]

Verify R4 (via PR #153) — full ensemble

Result: PASS-with-required-cleanup — R4 revert is clean,R1 work intact,F4 errata regression naturally resolved。 BUT 3-way drift surfaced(spec normative SHALL ≠ R4-softened CHANGELOG+rule ≠ reverted impl)+ rule line 195 self-inconsistency(Codex new finding)。
Full report: #153 (comment)

R5 cleanup loop required(small + targeted)。 User decision in master report。