[process·P1] Issue lifecycle and context hygiene: active queue ≠ open backlog

[atomchung]

Status

Implementation / process hardening — authority split recorded; one bounded docs leaf remains.

The lifecycle/context-loading contract and first stale-issue cleanup pass are complete. On 2026-07-30 the current M1 chain was restructured so routing, implementation, acceptance, roadmap, and future release history no longer share one issue body.

This work is non-blocking for M1. It must not interrupt #583 or #544.

Milestone

None — repository operating model.

Current product milestone remains M1. This issue owns context-loading efficiency, issue authority boundaries, and the bounded cleanup that proves them.

User outcome

A maintainer or coding agent can enter the repository, identify the current milestone and executable owner, and begin a bounded implementation without scanning the whole repository, replaying the open backlog, or relying on a host-specific instruction file for shared repository rules.

Current repository truth

Completed

• PR docs(process): an open issue is a record until the index says otherwise (refs #538) #541 landed docs/issue-lifecycle.md as the single lifecycle/context-loading authority, with thin routing pointers.
• The first bounded cleanup updated seven stale issue bodies/titles ([i18n·P2] zh-CN ships under a mixed-script waiver — parity ratchet, two missing copy groups, literal migration #387, [GTM·P2] Demo materials describe the step-② dialogue but never show an answer changing the card #46, [UX·P3] The share-safe card is documented but undiscoverable to the user (narrowed from 5 discoverability holes) #22, [test·discussion] SKILL-flow E2E acceptance: what is still owned here after ux_receipt and #486 #60, [P2·v2] v1/v2 session_id divergence can double-write one review's history (last open item of the four) #194, [Epic·superseded by #475] trade-review → 全生命週期個人投資 OS(找資訊→選股→交易→update→recap),分階段吸收 #12 and tracking: current context index — milestone, critical path, what an agent may pick up #27), marked [Epic·superseded by #475] trade-review → 全生命週期個人投資 OS(找資訊→選股→交易→update→recap),分階段吸收 #12 superseded, preserved history, and deliberately did not audit the long-tail backlog.
• tracking: current context index — milestone, critical path, what an agent may pick up #27 has now been compacted to execution routing only. It points to [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475, [bug·data·M1-blocker] Split-adjusted shares can be valued against prices on a different split basis #583, [process·qa] /fomo-qa needs one continuous host campaign across review, refresh, and consider #544, [QA·M0] Owner-live current-book and delivery acceptance on Claude and Codex #486 and [QA·M1] Owner-live TradeEvaluation acceptance #488 and no longer carries a merged-PR changelog, copied roadmap, detailed implementation design, or adjacent backlog narrative.
• [bug·data·M1-blocker] Split-adjusted shares can be valued against prices on a different split basis #583 now exposes a compact Implementation handoff for the active split-basis invariant, production entries, expected surfaces, required proof, non-goals, and stop condition.
• [process·qa] /fomo-qa needs one continuous host campaign across review, refresh, and consider #544 now exposes a compact Implementation handoff and is accurately marked blocked until [bug·data·M1-blocker] Split-adjusted shares can be valued against prices on a different split basis #583 stabilizes the arithmetic contract.
• [process·docs] Shared maintainer contract must not live in CLAUDE.md #587 now owns the separate documentation-authority defect: shared maintainer guidance lives in CLAUDE.md, causing Claude Code to treat a host-specific adapter as the preferred place for repository-wide rules.

Still true

• The long-tail backlog remains intentionally large. That is acceptable only while it stays outside the default execution capsule.
• Truth-critical changes still require targeted caller/route inventory because the codebase has real contract fan-out. “Do not scan the whole repository” does not mean “trust one issue body without searching affected production paths.”
• CLAUDE.md is currently 122 lines and mostly host-neutral maintainer material; [process·docs] Shared maintainer contract must not live in CLAUDE.md #587 is not yet implemented.

Decision

1. Separate four information lifecycles

Information	Authority	Must not contain
Product roadmap and scope decisions	#475 while it remains owner-ratified	Per-PR history, implementation instructions, active-queue bookkeeping
Implementation handoff and current evidence	Owning implementation issue and active PR	Other milestones, release marketing, unrelated completed work
Acceptance ledger	#486 / #488 and their eligible evidence	Implementation design owned elsewhere, speculative roadmap
Execution routing	#27	Release history, detailed design, copied roadmap, full dependency narrative

Release notes belong in a GitHub Release and/or CHANGELOG.md when an actual version/tag is cut. They do not belong in the default implementation context.

2. Use a bounded execution capsule

For substantive work, load and stop when sufficient:

1. latest main and AGENTS.md;
2. the directly affected current contract/reference;
3. tracking: current context index — milestone, critical path, what an agent may pick up #27 for current routing;
4. the owning implementation issue;
5. its direct acceptance gate;
6. the current PR, when one exists.

Search wider only for a named symbol, production-route inventory, duplicate, prior decision, or direct dependency. A repository-wide architecture review is a separate task.

3. Split by invariant or user outcome, not file

#583 remains one issue across price_feed, ValuationFrame, PortfolioBasis, review, both consider routes, and revisit because those are consumers of one price/share-basis invariant. Splitting by file would make partial fixes and semantic drift more likely.

#544 remains separate because it owns a different result: proof that the TradeEvaluation challenge actually reached the user.

4. Host-specific files are adapters, not shared authorities

#587 owns the correction:

• shared repository-maintenance guidance moves to a host-neutral maintainer guide;
• CLAUDE.md becomes a thin adapter containing pointers and actual Claude Code-only tooling behavior;
• AGENTS.md remains a thin cross-agent router;
• a small structural gate prevents CLAUDE.md from regrowing shared test, privacy, PR, architecture, or mirrored-surface sections.

A new rule may enter CLAUDE.md only when its owning issue/PR names the Claude-only mechanism it supports. “Claude may benefit from seeing this” is not sufficient.

Remaining scope

• Implement [process·docs] Shared maintainer contract must not live in CLAUDE.md #587 as one small docs/tooling PR, parallel-safe with M1 product work.
• Verify a fresh agent can identify and start [bug·data·M1-blocker] Split-adjusted shares can be valued against prices on a different split basis #583 or, after it closes, [process·qa] /fomo-qa needs one continuous host campaign across review, refresh, and consider #544 without loading unrelated issues or the full roadmap.
• Record PR/commit/test evidence on the relevant owning issue.
• Stop after these bounded checks. Do not audit the remaining backlog or add a project-management system in the same pass.

Non-goals

• Closing every old issue.
• Treating is:open as the execution queue.
• Splitting a coherent invariant into one issue per file/helper/test.
• Using one issue as release note + implementation summary + roadmap.
• Expanding AGENTS.md or CLAUDE.md into another complete project database.
• Adding GitHub Projects, complex lifecycle bots, or a generated roadmap before a repeated measurable failure requires them.
• Changing M1 scope or product behavior through process cleanup.

Acceptance criteria

• tracking: current context index — milestone, critical path, what an agent may pick up #27 contains only current execution-routing facts and links.
• Current implementation owners expose executable handoffs without requiring backlog replay.
• Roadmap, implementation, acceptance, and release-history facts have distinct authorities.
• Normal development does not load all open issues or scan unrelated repository areas.
• Cross-cutting correctness work still performs targeted production-route/caller inventory.
• [process·docs] Shared maintainer contract must not live in CLAUDE.md #587 moves shared maintainer guidance out of CLAUDE.md, leaves a thin Claude adapter, and adds a bounded regression gate.
• A fresh maintainer can name the milestone, active front, acceptance gate, contention, next step, and stop boundary from the bounded capsule.
• M1 implementation is not delayed by this process work.

Dependencies

• tracking: current context index — milestone, critical path, what an agent may pick up #27 — current execution routing; owns no implementation, roadmap, acceptance, or release history.
• [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475 — product roadmap/scope guard.
• [bug·data·M1-blocker] Split-adjusted shares can be valued against prices on a different split basis #583 / [process·qa] /fomo-qa needs one continuous host campaign across review, refresh, and consider #544 — current and next M1 implementation owners.
• [QA·M0] Owner-live current-book and delivery acceptance on Claude and Codex #486 / [QA·M1] Owner-live TradeEvaluation acceptance #488 — M1 acceptance owners.
• [process·docs] Shared maintainer contract must not live in CLAUDE.md #587 — shared maintainer authority / thin CLAUDE.md adapter implementation.
• [research·process] Development operating model: measure rework, gate cost, and multi-agent contention before the next capability wave #483 — broader operating-model research.
• PR docs(process): batch tests and isolate edge-case tracks #585 — parallel docs-only batching/edge-case guideline.

Privacy boundary

Use only public repository metadata, public issues/PRs, synthetic fixtures, and generic examples. No real trade, holding, amount, date, thesis, source path, private-repository content, or private QA text may enter process evidence.

Rollout / compatibility / recovery

Issue-body edits and documentation moves are reversible through GitHub history. No runtime, schema, or user data change is owned here. If a thinner routing surface omits a necessary pointer, restore the pointer at its authority; do not restore copied history or design prose.

Historical first-pass result

The 2026-07-29 pass updated seven issue bodies/titles, marked #12 superseded, closed no issue, preserved historical bodies, and landed PR #541 with the lifecycle authority. Full per-candidate evidence remains in this issue's comments and GitHub revision history.

[atomchung]

Initial setup evidence — 2026-07-29

This issue is now the lifecycle/context-hygiene owner.

Completed in this setup pass:

• Searched open issues for an existing dedicated owner; only tracking: current context index — milestone, critical path, what an agent may pick up #27 contained the audit, and it was a routing index rather than an implementation owner.
• Created this bounded process issue with the lifecycle contract, default context-loading order, initial six-candidate triage set, non-goals and acceptance criteria.
• Updated tracking: current context index — milestone, critical path, what an agent may pick up #27 so it is now a small current context index and links [process·P1] Issue lifecycle and context hygiene: active queue ≠ open backlog #538 as the cleanup owner.
• tracking: current context index — milestone, critical path, what an agent may pick up #27 now states the product critical path separately from this non-blocking process lane: [process·qa] /fomo-qa has no verified walkthrough for snapshot_review, the route M0-U02 starts on #526 → [QA·M0] Owner-live current-book and delivery acceptance on Claude and Codex #486 for M0; [process·P1] Issue lifecycle and context hygiene: active queue ≠ open backlog #538 for bounded issue hygiene.

Not done yet:

• No stale candidate was closed or materially rewritten in this setup pass.
• [Epic·superseded by #475] trade-review → 全生命週期個人投資 OS(找資訊→選股→交易→update→recap),分階段吸收 #12 has not been closed or formally dispositioned; that remains an owner-approval boundary.
• docs/issue-lifecycle.md / AGENTS.md / development-guide pointers have not been implemented yet.

Next safe step under this issue:

1. Revalidate and update only [i18n·P2] zh-CN ships under a mixed-script waiver — parity ratchet, two missing copy groups, literal migration #387, [GTM·P2] Demo materials describe the step-② dialogue but never show an answer changing the card #46, [UX·P3] The share-safe card is documented but undiscoverable to the user (narrowed from 5 discoverability holes) #22, [test·discussion] SKILL-flow E2E acceptance: what is still owned here after ux_receipt and #486 #60 and [P2·v2] v1/v2 session_id divergence can double-write one review's history (last open item of the four) #194.
2. Add a non-destructive Superseded by #475 status header to [Epic·superseded by #475] trade-review → 全生命週期個人投資 OS(找資訊→選股→交易→update→recap),分階段吸收 #12, but do not close it without explicit owner approval.
3. Stop after that bounded set and report whether tracking: current context index — milestone, critical path, what an agent may pick up #27 is sufficient for a fresh agent to identify the current M0 next step.

Do not open additional cleanup issues during this pass unless one candidate contains genuinely distinct residual work with no existing owner.

[atomchung]

Claiming the bounded first pass in this session (2026-07-29): #387 / #46 / #22 / #60 / #194 body revalidation, a non-destructive supersession header on #12, #27 refresh to latest main, and the docs/issue-lifecycle.md authority plus its two thin pointers. No issue will be closed in this pass. Evidence comment to follow.

[atomchung]

First bounded pass — complete. 2026-07-29, against main@889c8fd

Every disposition below was re-derived from current main, not from the issue thread. Where a body's claim survived, it survived because the code or the README still shows it.

Reported numbers

• Bodies updated: 7 — [i18n·P2] zh-CN ships under a mixed-script waiver — parity ratchet, two missing copy groups, literal migration #387, [GTM·P2] Demo materials describe the step-② dialogue but never show an answer changing the card #46, [UX·P3] The share-safe card is documented but undiscoverable to the user (narrowed from 5 discoverability holes) #22, [test·discussion] SKILL-flow E2E acceptance: what is still owned here after ux_receipt and #486 #60, [P2·v2] v1/v2 session_id divergence can double-write one review's history (last open item of the four) #194, [Epic·superseded by #475] trade-review → 全生命週期個人投資 OS(找資訊→選股→交易→update→recap),分階段吸收 #12, tracking: current context index — milestone, critical path, what an agent may pick up #27.
• Titles corrected: 7. Not in the original plan, but a title is what gh issue list shows first, so a [P1] on a P3 is the same defect this issue exists to fix. [UX·P3] The share-safe card is documented but undiscoverable to the user (narrowed from 5 discoverability holes) #22 P1→P3, [GTM·P2] Demo materials describe the step-② dialogue but never show an answer changing the card #46 P1→P2, and four titles that named work already shipped.
• Closed: 0. By design — this pass closes nothing.
• Marked superseded: 1 ([Epic·superseded by #475] trade-review → 全生命週期個人投資 OS(找資訊→選股→交易→update→recap),分階段吸收 #12, non-destructively; still open, still needs owner approval to close).
• Historical bodies preserved: 6, each inside a <details> block whose summary states why it is no longer accurate.

Per-candidate disposition

Issue	Was	Now	Evidence that changed the verdict
#194	4 hardening items	1 item	Items 1/2/4/5 shipped (tests/test_digit_ban.py, PR #209, PR #212). Item 3 confirmed still live: review.py:3031 hashes f"{nonce}|{route}|{language}", coach.py:145,154 hashes the bare nonce.
#22	5 P1 discoverability holes	1 P3 hole	Holes 1, 2 and 5 closed; 4 substantially closed. Hole 5's fix carries a code comment citing #22.5 — the work was done and the issue never told.
#46	"differentiator invisible everywhere"	one missing artifact	README now states the step-② claim in three places. What is genuinely absent is the before/after: the same position, two answers, two cards.
#387	RFC to add zh-CN	residual purity work, reordered	zh-CN has been live since PR #391. Confirmed today: zh-CN.json has 75 top-level keys vs 83, missing condition_state and condition_check, and load_copy has no cross-locale fallback — so condition_state_line silently renders nothing for every zh-CN session.
#60	"no E2E test exists"	one scoping question	All four named symptoms shipped; the residual is mostly re-homed.
#12	live epic premise	Superseded by #475	#475 explicitly reverses the full-investment-OS premise.

Residuals that gained an owner

• [P2·v2] v1/v2 session_id divergence can double-write one review's history (last open item of the four) #194 → the session_id divergence is now stated with both call sites and line numbers; someone can start from the body.
• [i18n·P2] zh-CN ships under a mixed-script waiver — parity ratchet, two missing copy groups, literal migration #387 → four ordered items, with "fix the ratchet before the keys" made explicit. test_copy_ratchet.py diffs en vs zh-TW only and persona_sweep.LOCALES is ("zh-TW", "en") — that blind spot is why the missing groups shipped green, so hand-filling the keys first would fix one instance and leave the generator intact.
• [UX·P3] The share-safe card is documented but undiscoverable to the user (narrowed from 5 discoverability holes) #22 → the residual is named and its original prescription is marked superseded: it asked for a hint line appended to the private card, which the owner has since ruled against. Without that note the next agent implements a rejected fix.
• [test·discussion] SKILL-flow E2E acceptance: what is still owned here after ux_receipt and #486 #60 → surfaced as needing a ruling rather than a PR. Its four residuals: two unclaimed and narrow, one re-homed to [eval·P2] The rubric judge has never been run — calibrate it before its score means anything #454, one architecturally ruled out by [design·P1] eval 的對象是「判斷的交派紀律」不是判斷正確性:三層判斷 × 現有 eval 資產盤點 #159 and answered by ux_receipt + [QA·M0] Owner-live current-book and delivery acceptance on Claude and Codex #486. Recommend: rewrite to own only the two unclaimed items, or close as superseded and re-open them as one small leaf. Owner's call.

#27 — re-ranked against the milestone, and one correction

The previous version listed #526 as a ready-to-execute blocker. It had already merged in #537 a few minutes earlier. It also recorded that no CI evidence was available for the head commit; the tests workflow shows success on 889c8fd.

The re-rank is the substantive part. With #526 closed, the M0 critical path collapses to a step no agent can perform — #486 owner-live acceptance. So #27 now separates:

1. the owner-only next step;
2. what an agent may pick up while that waits, ranked by likelihood of breaking the [QA·M0] Owner-live current-book and delivery acceptance on Claude and Codex #486 walk rather than by severity label — [bug·design·P2] Every new snapshot declaration re-asks for theses it was already given #539 first, because a second review that re-asks every thesis makes the product look memoryless at the exact moment memory is what is being accepted;
3. what needs a ruling before it is executable ([design·M0] The recorded book carries a provenance record the user can act on #532, [test·discussion] SKILL-flow E2E acceptance: what is still owned here after ux_receipt and #486 #60);
4. future-milestone and research records, excluded by default.

Contract landed

PR #541 — docs/issue-lifecycle.md as the single authority, plus a route pointer in AGENTS.md (shaped like the existing QA-route paragraph) and one line in docs/development-guide.md's complements list. Full offline suite 40/40.

No checker, deliberately: a check can verify a Status header exists, not that the stated status is true. A green suite over a backlog that still lies is worse than no check. The file records the condition under which one gets added.

Acceptance criterion — honest read

Can another agent start from #27 and correctly name the milestone, owning issue, blocker and next smallest step without replaying the backlog?

For the product path: yes. #27 names M0, #485 as implementation owner, #486 as the gate, and states that the gate is owner-only.

Caveat worth recording rather than glossing: that answer is only true while #27 is fresh. It went stale within hours last time — one merge did it. The index is a manual surface with no gate, which is the same class of dependency this repository has been burned by before. If the second pass finds it stale again, that is the signal to make freshness mechanical rather than to write a better index.

Not done — and why

• Nothing was closed. [Epic·superseded by #475] trade-review → 全生命週期個人投資 OS(找資訊→選股→交易→update→recap),分階段吸收 #12 needs owner approval; [test·discussion] SKILL-flow E2E acceptance: what is still owned here after ux_receipt and #486 #60 needs an owner disposition.
• The remaining ~90 open issues were not touched. Per this issue's own non-goals.
• No labels were added. Worth a decision before the second pass: the repository has 10 labels, none of them lifecycle, so status lives only in issue bodies and in this index. A small status:* label set would make is:open filterable and is the most direct answer to "too many issues" — but it is a second mechanism, and this issue's design says test the simple contract first. Recommend deciding it after measuring this pass, not now.

Stopping here, as instructed

The pass stops. Whether it helped is measured by the next agent that starts cold, not by this comment.

[atomchung]

Second pass — owner dispositions applied. 2026-07-29

The first pass deliberately closed nothing. The owner then ruled on the two items it had escalated, and authorized closing work that is genuinely no longer necessary. Open issues: 95 → 87.

Every close below carries its evidence in the issue's own closing comment. None was closed for age.

#60 — closed, replaced by #542

Owner picked the "close and re-open the residual as one small leaf" option. #542 owns what actually remains: B-1's whitewash-label assertion and B-9's upgrade from a whole-card number proxy to the section-scoped rule it stands in for — both verified live in washer.yaml's pending: block and check_card.py:395-397.

The strongest argument for closing rather than narrowing again: the AskUserQuestion headless ceiling was ruled out of automated scope by #159, and an open issue titled "no E2E acceptance test" is a standing invitation to rebuild what was deliberately not built. #542 states that exclusion explicitly.

Closed as out of scope — the roadmap killed them, not time

Issue	Why
#35 Not-Buy Tracker	#475's out-of-scope list names it verbatim: automatic tracking of every trade considered but not taken. The bounded version already exists as consider --resolve.
#38 KOL/research source attribution	Its design is inference-first over a D-14 KOL corpus — that corpus is #475's KOL platform exclusion. The surviving half (user-supplied evidence references) is #479.
#34 Behavioral heuristic warnings	Its own 2026-07-08 comment already split it: post-hoc absorbed by #146, pre-trade form is now M1 #479. Carried forward the part that would have been lost — the tags #146 declared still have no writer, which is #455.
#6 user-stories.md calibration	The file it edits is not in this repository. Its content is settled by #475 and docs/decision-fomo-kernel-shape.md.

Closed as delivered

Issue	Evidence
#37 Market context	Context version PR #145; the numeric residual it routed to #154 landed as TWR-vs-market (PR #182) and account IRR (PR #179). Verified #154's closing ledger rather than assuming.
#199 Thesis-reconciliation discovery	Its own gate ("if the #412 rulings satisfy this, it can close") was met by #430/#433/#437/#438 and #416's ruling. Confirmed and closed.
#262 Card feedback: 完全沒有感覺	Symptom ① → #305 closed; ② → #279/#282 shipped plus #356's ruling; ③ → #304 closed, residual is #230. The subjective half belongs to #486, which is literally a first-review owner walk on latest main.

Kept, but no longer readable as active work

Three records were mislabelled rather than obsolete. Each got a status header and a corrected title:

• [P2·deferred] revenge_trade + overtrading 落到 engine — 排在 #455 之後 #5 revenge_trade / overtrading → Deferred behind [bug·design·P2] exit_anxiety / horizon_break / fomo_entry have copy and a declared home but no writer — and horizon_break's missing opportunity flag makes any rule on it permanently skipped #455. Adding two more behavior tags while three declared ones have no writer is wrong-ordered, not wrong.
• [P2·parked] chat 引流版：把 skill 搬到純 chat 介面（#475 Phase 4 之後才解凍） #8 chat distribution → Parked. Its original unfreeze condition (Stage 0, [P0] Stage 0：把卡塞到一個真人面前驗「有沒有戳中」（需求側 0 驗證） #3) passed, but [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475 replaced the condition: adapters are opt-in and come after validation, never as a way to widen the funnel first.
• [parked] 支援「先賣後買 / 做空」：目前刻意 long-only 並誠實揭露 #15 short selling → Parked. Long-only is a deliberate decision with honest degradation (orphan_sells() discloses rather than miscomputes). What would unpark it is a real statement containing short rows, not appetite.

What this pass deliberately did not do

No sweep of the remaining ~87. The ones closed here all had a specific piece of evidence — an explicit #475 exclusion, a closed replacement issue, or a shipped PR named in their own thread. Everything without that evidence stayed open, including several that look old.

docs/issue-lifecycle.md gets no amendment from this pass; the rules it states held.

Both passes are now complete. Measurement — whether a cold-start agent can name the milestone, owner, blocker and next step from #27 alone — is the gate on anything further.

[atomchung]

Leaf #587 shipped: PR #593, squashed to main@78aedc4. Recording it here because #587's acceptance criteria asked for it, and because this leaf changes where a contributor's context comes from.

What landed. AGENTS.md is now the shared always-on instruction floor for every client — route selection, non-negotiable boundaries, and an instruction-authority policy stating that file-load order never lets a host adapter change product truth. docs/maintainer-guide.md is the new detailed host-neutral maintainer contract, holding the tests, privacy boundary, PR and worktree rules, and the mirrored-surfaces map. CLAUDE.md went from 51 KB to 1.7 KB: the @AGENTS.md import, its own edit boundary, and the one genuinely Claude-only mechanism.

Relevance to this issue's context-hygiene mandate. Two effects, opposite in sign:

• Claude sessions stop paying ~51 KB of always-on context on every task, including ones that only touch documentation.
• The mirrored-surfaces map moves from always-on (for Claude) to routed (for both clients). docs/development-guide.md §6 is explicit that a soft-routed document may never be read, so this is content-delivery traded for obligation-delivery. AGENTS.md states the read obligation at boundary strength, and nearly every row of the map already names its own drift test, so the map is a navigation aid rather than the enforcement — but the trade is real and stated rather than glossed.

Verified, not assumed. A fresh headless Claude Code 2.1.220 session with no tools quotes AGENTS.md's authority policy verbatim with the import present, and answers NOT LOADED with the import removed. So the import is load-bearing, this version does not discover AGENTS.md on its own, and it is not a double load. The smoke is documented in the new guide so the chain can be re-checked on any client.

Also fixed in passing: 34 citations across 24 files named CLAUDE.md as the home of a rule that moved. All retargeted, with a gate that keeps them pointed at the right file — three were found only because a truncated grep was re-run without the truncation.

No product change, proven rather than asserted: an AST comparison across all 18 touched .py files shows executable structure changed only in the two test files; the one schema edit is inside a description. Full offline suite 46/46, CI green on 3.11 and 3.12.

Two follow-ups opened, both discussion records rather than blockers:

• [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 — CI runs the suite on every push for every client, but it reports; main has no required status check, so the only gate that blocks is the Claude Code hook. Closing that is a repository setting plus a policy call.
• [process·docs] Nothing checks that a gate the mirrored-surfaces table names still exists #594 — nothing checks that a gate the mirrored-surfaces table names still exists. Scoped deliberately narrow: it adds the missing check and changes no prose. The wider idea (replace the rows with pointers at their gates) is recorded there as rejected, with reasons, so it is not re-proposed.

[atomchung]

Bounded backlog-truth audit — 2026-07-30

The M1 cleanup scan against main@dabf709 found that the lifecycle mechanism is correct, but several open issue bodies still carry obsolete execution claims. This is recorded here as the owner of issue hygiene; none should interrupt #612/#618.

Corrected in this pass

• tracking: current context index — milestone, critical path, what an agent may pick up #27 — was still routing to [design·M1] Decision-first TradeEvaluation answer — one supported judgment, not a disclosure dump #579/[architecture·implementation] One market-data resolver — minimum Yahoo calls, one fact bundle for prepare and consider #605 after PRs docs(contract): a consider answer leads with one supported decision tension, not a disclosure dump (closes #579) #608/feat(market-data): one resolver, one dated bundle, and a consider that can see today's prices (refs #605) (closes #584) #611 had landed and claimed no open PRs despite docs(registry): profile.md has no writer, and both surfaces that claimed one now say so (refs #58) #614/test(consider): the canonical-book refusal fails closed on both branches that reach it (closes #613) #615. Rewritten to the current two-leaf queue ([defect·P1] prepare aggregates a missing FX rate as 1.0 — #602's refusal never reached the review lane #612 → [implementation·M1] A consider answer states which day's prices it was computed from #618 → [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610).
• [defect·P1] prepare aggregates a missing FX rate as 1.0 — #602's refusal never reached the review lane #612 — body still presented the fail/degrade choice as undecided after the owner had ruled refuse. Rewritten as an executable handoff with compatibility and recovery criteria.
• fix(card): pre-commitment preview falsely implies a standing rule exists on a user's first-ever review #546 — lacked lifecycle headers and current evidence. Reverified on latest card_renderer.py, marked valid but adjacent/non-M1, and given a bounded acceptance contract.

Still stale; next bounded hygiene pass only

• [feat·M1] Bounded TradeEvaluation with visible two-sided challenge #479 — its header still says [process·qa] /fomo-qa needs one continuous host campaign across review, refresh, and consider #544 receipt work is missing and names [process·qa] /fomo-qa needs one continuous host campaign across review, refresh, and consider #544 as the only active front, while [process·qa] /fomo-qa needs one continuous host campaign across review, refresh, and consider #544 is closed and [QA·M1] Owner-live TradeEvaluation acceptance #488 explicitly says the formal acceptance walk is deferred/non-blocking. The implementation result remains valid; only the current-summary/remaining-scope section is stale.
• feat: 第一次復盤後從行為推斷 profile.md + theses.jsonl（少問、猜為主） #58 — still reads as an executable profile.md writer design. [research·product] Philosophy bootstrap: decision guidance before transaction history #597 rejected that file as a second canonical truth, [decision] profile.md 是記憶體系第 2 層的化石:退役、復活,還是等 #446 刀 2 搬家? #619 now owns retire/revive/move, and [design·P1] The distillation layer: bounded profile memory over the append-only ledger (scoping) #446 owns the structured distillation successor. It needs a Superseded/decision-routing header rather than implementation language.
• [feat·data·M0] Canonical recorded book + low-friction snapshot refresh #485 — its sequencing header still says finish [arch·data·M0] No-write virtual review book + pre-append state-version gate #501 then execute the older recorded-book chain. The capability has moved far beyond that sequence; the body is useful history but not a current handoff.

Decision

Do not bulk-rewrite the rest of the backlog. After the current product leaves stop, run one small hygiene pass limited to the three bodies above, preserving historical content under <details>. No new process checker is justified: this is stale truth, not missing header syntax.