knowledge: 2 insights — park a transient async state behind a test-controlled gate, a PASS body that records a surviving mutant is a finding (1 new page, 2 amended, 26 plan-gaps retired)

[choiyounggi]

Knowledge flush — 2 insight(s) + 26 plan-gap rows

Flush run id 20260928-092537-1352 (inherited from the auto-flush hook that spawned this session via DEV_LOOP_FLUSH_RUN_ID; step-0 acquire returned already-owned — a first attempt with a fresh run id was refused with held 20260928-092537-1352 81s, and the process tree showed the holder is this session's own parent hooks/auto-flush.sh). Branch knowledge/choiyounggi-20260928-092831 off origin/main at 5986311. Claimed via queue-claim.js claim → 28 ids: 2 session candidates (f6e9703c7f44b032, 3693d1c0d6feb6d3) + 26 plan-gaps.jsonl rows.

Outcome: 1 new page, 2 amended pages, 2 domain indexes updated, no new category; 2 of 2 session candidates ingested; 26 of 26 plan-gap rows excluded as project-specific (Local-layer candidates below).

Gates run on the edited checkout: node scripts/wiki-structure-checks.js wiki → pages: 354, indexes: 13, findings: 0 (untouched baseline: 353 / 0); node scripts/wiki-lint-prohibitions.js → violations: 0 (baseline 0); new page body 76 lines; no banned qualifier in any added line (grep for usually/often/consider/might want/generally/typically/probably/as appropriate over git diff -U0 -- wiki/ + the new page → 0 hits after two rewrites).

Verified best-practice

1. Park an async orchestrator behind a test-controlled decision to assert on a transient mid-run state (f6e9703c7f44b032, repo t3-swap-flake)

• Claim. When a test must read a mid-run state that a background finisher/reaper will wipe, drive the system into a state whose only exit is a decision the test controls (unresolved approval gate, withheld ack), assert inside that park, then release and await the finisher. Reuse the gate's own fixture; disable every other exit (tick timeouts); prove by mutating every finisher spawn site and looping.
• Sources checked (fetched 2026-09-28).
  • https://martinfowler.com/articles/nonDeterminism.html — "Never use bare sleeps to wait for asynchonous responses: use a callback or polling"; Humble Object so logic "can be tested synchronously". The park is the synchronous-ordering form of that advice for a state that is transient (polling cannot distinguish "not yet" from "already wiped").
  • https://testing.googleblog.com/2021/03/test-flakiness-one-of-main-challenges.html — "You should never need any type of sleep or wait in your test automation … Wait() is as good as Hope()".
  • https://tokio.rs/tokio/topics/testing — paused time makes timer-driven futures deterministic ("any time-related future may become ready early"); it orders timers, not tasks against each other — recorded as an edge row so the page does not overclaim what paused time buys.
  • https://docs.rs/tokio/latest/tokio/task/struct.JoinHandle.html — destructor-before-await-returns guarantee: the handle is the ordering point for the post-wipe state; the park is for the pre-wipe state.
• Local reproduction (this flush, Node v26.7.0, script written under the project's .claude/tmp/ and deleted in the same command): a startup seeds a Map and spawns a finisher that clears it on the next timer turn. Read right after startup → 0/200 failures (looks deterministic). Same read after one scheduler turn of pressure → 200/200 failures. Finisher parked behind a promise the test resolves after the read, same pressure → 0/200 failures and the map is empty after release every run.
• Field evidence (recorded by the originating session, not re-run here): escalation_timeout_ms = 0 disabled the lead's tick branch (crates/crew-lead/src/dispatch.rs:576), so a planted violation with max_rework = 0 held the lead task open until resolve_gate; mutating both tokio::spawn(reap_controls_entry( sites → FAILED ... Elapsed(()) at 30.03 s; restored → ok. 4 passed; 20/20 stable.
• Confidence: verified (official-doc principles + a reproducible local check + field mutation evidence).

2. A PASS verdict whose body records a surviving mutant is a finding, not a clearance (3693d1c0d6feb6d3, repo t5-proto)

• Claim. When a test-quality auditor or reviewer returns PASS but its body notes that a mutation survived (labelled "secondary"), map every surviving mutant to the brief's own requirement list; a guarantee the brief states literally is a missing test whatever label the reviewer gave it — add the killing case and resume the same reviewer to reproduce the kill.
• Sources checked (fetched 2026-09-28).
  • https://pitest.org/quickstart/basic_concepts/ — "Survived: the mutation was not detected by the covering test" (verbatim re-fetched); the survivor is a coverage statement independent of any reviewer's verdict scope.
  • https://mutants.rs/using-results.html — "missed — No test failed with this mutation applied, which seems to indicate a gap in test coverage" (verbatim).
  • Local reading of this repo's agents/test-quality-auditor.md: the FAIL list is no-assertion / disabled cases / uncovered changed behavior / rubber-stamping / the case-count floor, output is VERDICT: PASS | FAIL + one justification line — coverage of every stated guarantee is not among the FAIL conditions, so a PASS body can carry a surviving mutant without contradiction. This is the mechanism the candidate's "why" described.
  • Already-cited on the merge target: https://google.github.io/eng-practices/review/reviewer/standard.html (reviewers work from partial context; facts override opinion).
• Field evidence (recorded by the originating session, not re-run here): cargo test -p crew-proto 12→13 tests; under the deny_unknown_fields mutant unknown_entry_fields_are_tolerated ... FAILED; under a .rev() mutant malformed_entry_is_dropped_and_valid_sibling_kept ... FAILED; restored → 82 passed; the resumed auditor reproduced both kills and returned PASS.
• Confidence: the survivor semantics are verified against the tool docs; the review-process directive is field-tested (one production incident + the auditor contract). Merged into pages that already carry field-tested (evaluating-review-feedback) and verified (surviving-mutant-equivalence-triage, whose new row is a routing pointer plus field evidence, not a new mechanism claim); neither page's confidence was changed.

Existing-layer check

Pages read: testing-async-async-testing, testing-flaky-diagnosing-flaky-tests, testing-quality-tests-that-cannot-fail, testing-quality-narration-based-ordering-assertions, testing-quality-proving-a-critical-section-is-lock-protected, testing-quality-surviving-mutant-equivalence-triage, testing-quality-harness-reverse-controls, testing-quality-minimum-case-set, qa-process-evaluating-review-feedback, qa-process-fresh-context-code-review, qa-process-completion-claims, testing-async-transient-state-behind-a-controlled-gate

Semantic dedupe (wiki_search, k=5) top hits — candidate 1: infrastructure-agent-orchestration-control-signals-vs-primary-artifacts (0.729), testing-async-async-testing ×2 (0.728, 0.725), qa-environments-test-environment-parity (0.722), testing-data-testcontainers-reaper-on-docker-desktop-macos (0.720). Candidate 2: testing-quality-tests-that-cannot-fail ×2 (0.764, 0.762), testing-quality-narration-based-ordering-assertions (0.758), qa-process-completion-claims (0.758), testing-quality-unasserted-return-fields (0.747). No hit's "When this applies" describes either candidate's situation.

Candidate	Overlap found	Decision
1 park	testing-async-async-testing covers waiting for a final state (condition wait, fake timers, completion handle) — the transient-state case is absent; testing-flaky-diagnosing-flaky-tests sleep row says "wait on the completion condition", which a transient state has no form of. No conflict.	New page wiki/testing/async/transient-state-behind-a-controlled-gate.md. Back-link: async-testing gains a table row + related:; testing/index.md async section gains the row. related: one-way to diagnosing-flaky-tests, tests-that-cannot-fail, narration-based-ordering-assertions, proving-a-critical-section-is-lock-protected. Reciprocal related: on diagnosing-flaky-tests and tests-that-cannot-fail deliberately deferred: open PR #223 rewrites both pages' related: lines, and a second edit here would conflict on merge (see Open-PR check).
2 PASS+survivor	testing-quality-surviving-mutant-equivalence-triage classifies a survivor on code you own but has no row for "reported inside a passing review"; qa-process-evaluating-review-feedback is the page for "findings arrived, decide what to implement" and already carries the partial-context mechanism. qa-process-fresh-context-code-review step 4 ("return only the verdict") is not a conflict — it concerns what flows back into the producing session's next round, while this candidate reads the reviewer's own report body.	Merge into evaluating-review-feedback (trigger sentence, edge row, instead-of row, 2 sources + local reading + field evidence, related: +surviving-mutant-equivalence-triage +fresh-context-code-review) and a reciprocal edge row + field evidence + related: on surviving-mutant-equivalence-triage; qa/index.md load-when widened.

Conflicts flagged: none.

Open-PR check

gh pr list --repo choiyounggi/dev-loop --state open --search "head:knowledge/" → two open heads: #223 knowledge/choiyounggi-20260927-220735 and #225 knowledge/choiyounggi-20260928-082803. Both fetched; git diff --name-only origin/main origin/<head> -- wiki/ read for each, plus the full diff hunks on the pages nearest to each candidate.

Head	Files touching either candidate's area	Overlap	Verdict
#223	wiki/testing/flaky/diagnosing-flaky-tests.md (+row: when the test holds the finisher's JoinHandle, await it between write and read to force the bad interleaving red, before the write for green; +edge row on stress-loop reproduction rates; +tokio JoinHandle source), wiki/testing/mocking/fake-server-forward-before-reply.md (new), tests-that-cannot-fail.md + captured-call-arguments.md (related: only)	Complementary, not duplicate: #223's row forces a reproduction of the race via a completion handle the test already owns; candidate 1 is the case where the pre-wipe state must be observed green and no handle helps because the state must still be present — the new page's "no external decision point" row routes to #223's technique, and the JoinHandle guarantee is cited as the ordering point for the post-wipe state. No text on candidate 2's topic.	candidate 1 new; candidate 2 new
#225	wiki/testing/quality/alias-table-contract-tests.md, wiki/testing/mocking/fake-intersection-observer-for-viewport-animations.md (new), wiki/testing/index.md (rows in quality/mocking)	Its only mutation-related text is a "hardcode the defaults and require red" step and PIT/Google citations on alias tables — a different trigger from "PASS body records a survivor". Nothing on transient async state.	candidate 1 new; candidate 2 new
#223 / #225	26 plan-gap rows	none (disjoint hash set from #223's 100 and #225's 77 retired rows)	drop — project-specific (below), not pending-duplicate

Merge-conflict note for the owner: this PR and #223 both add a row to wiki/testing/index.md (different sections) and #223 also touches wiki/qa/index.md on a different line; diagnosing-flaky-tests.md and tests-that-cannot-fail.md were left untouched here for that reason.

Routing decision

Candidate	Layer	Domain/category	Page	Why here
1 park	bundled	testing / async	new testing-async-transient-state-behind-a-controlled-gate	The directive is about how to test an async ordering deterministically, the async category's charter; flaky is for a test that is already intermittent, and this page is the design that keeps it from becoming one. No new category.
2 PASS+survivor	bundled	qa / process (+ testing / quality pointer)	merge qa-process-evaluating-review-feedback; reciprocal row on testing-quality-surviving-mutant-equivalence-triage	The trigger is "a review verdict arrived and you are deciding what it means", i.e. the process page; the testing page owns what to do with the survivor once you accept it is one. No new category.

Local-layer candidates

All 26 plan-gaps.jsonl rows are wiki-plan Phase B decisions of the form "Planning tN: deciding X (no owning wiki page)" whose directives name one repository's own files, symbols, values, or route tables (Kotlin LinklyTextStyle/LinklyColor alias tables, ScreenContainer.kt, dispatch.rs :520-537, mock-source.ts:445, chat.css selectors). They fail the layer test (wiki-ingest step 3: would be wrong in another codebase) and are excluded from this PR; run wiki-ingest inside the named project to file them. Disposition recorded on each retired row: dropped: project-specific plan-gap.

Hash	Project (cwd)	Decision	wiki-local target
1bfe8ff35d7ba96d	dace (~/Desktop/workspace/linkly-calendar/dace)	t5-screen-restyle: full legacy-name grep-to-zero set	wiki-local/frontend/design-system/legacy-token-grep-to-zero-set.md
8085b35b918b9ccc	dace	t5-screen-restyle: alias-to-role rename mapping	wiki-local/frontend/design-system/alias-to-role-rename-mapping.md
cfa1be9fd8184e59	dace	t5-screen-restyle: AppRoot.kt's 2 legacy references	wiki-local/frontend/design-system/approot-legacy-reference-rename.md
fac0e3b1760305fa	dace	t5-screen-restyle: floating tab dock geometry	wiki-local/frontend/design-system/floating-tab-dock-geometry.md
86daa209438f656c	dace	t5-screen-restyle: per-tab-item pill styling	wiki-local/frontend/design-system/tab-item-pill-styling.md
4d3301ae5c046220	dace	t5-screen-restyle: tab-hosted-screen bottom inset	wiki-local/frontend/design-system/tab-hosted-screen-bottom-inset.md
2c6aa03262ad25b3	dace	t5-screen-restyle: tab-hosted-screen bottom inset (corrected against verified call sites)	same page as 4d3301ae5c046220 — the corrected revision supersedes it
ab0c5d54ca599d04	dace	t5-screen-restyle: ChatScreen composer inset	wiki-local/frontend/design-system/chat-composer-bottom-inset.md
ecf4bac7d47b8f02	dace	t5-screen-restyle: MainTabScaffold structural change	wiki-local/frontend/design-system/main-tab-scaffold-overlay-dock.md
c94ce34b34a9aff9	dace	t5-screen-restyle: LinklyChatBubble v2 shape	wiki-local/frontend/design-system/chat-bubble-v2-shape.md
9d255cdb6b4cb743	dace	t5-screen-restyle: LinklyChatBubble partner avatar	wiki-local/frontend/design-system/chat-bubble-partner-avatar.md
cfb4d875e927de2d	dace	t5-screen-restyle: bubble timestamp text style	wiki-local/frontend/design-system/chat-bubble-timestamp-style.md
b48f6b37879ac956	dace	t5-screen-restyle: CoupleSetupScreen invite-avatar	wiki-local/frontend/design-system/couple-invite-avatar-initial.md
dd1fe7e62611c5d5	dace	t5-screen-restyle: SettingsScreen couple-avatar header visibility rule	wiki-local/frontend/design-system/couple-avatar-header-visibility.md
9d23ccad0e9090fa	dace	t5-screen-restyle: CoupleTogetherDaysCalculator	wiki-local/frontend/design-system/couple-together-days-calculator.md
dc85e01d2df70e0f	dace	t5-screen-restyle: CoupleAvatarDayHeader placement and content	wiki-local/frontend/design-system/couple-avatar-day-header.md
d7353fab12a4f0fb	dace	t5-screen-restyle: PhotosScreen scope	wiki-local/frontend/design-system/photos-screen-token-migration-scope.md
8e19059c33c9018e	dace	t5-screen-restyle: ChatImageViewer hardcoded literals	wiki-local/frontend/design-system/chat-image-viewer-scrim-tokens.md
54462465e1d06013	dace	t5-screen-restyle: ChatBannerCenter.kt (no change, verified carve-out)	wiki-local/frontend/design-system/chat-banner-center-carve-out.md
80edbc00c185b6f3	handfish (~/Desktop/workspace/linkly-crew/handfish)	t6-lead-artifacts-ondisk: same-name entries; dedup across DodCheck::Artifact and artifacts_expected	wiki-local/backend/lead-judge/artifact-name-dedup-across-dod-and-expected.md
5e8b85977b985e26	handfish	t6-lead-artifacts-ondisk: how dispatch passes the cwd	wiki-local/backend/lead-judge/role-cwd-resolved-once-in-on-envelope.md
2c8dbcd839163f67	handfish	t6-lead-artifacts-ondisk: sync vs async filesystem calls	wiki-local/backend/lead-judge/sync-std-fs-in-judge.md
0f78e800933d2aaa	handfish	t6-lead-artifacts-ondisk: symlink and special-file fixtures	wiki-local/backend/lead-judge/symlink-and-fifo-fixtures.md
1a92fa259582f25b	handfish	t8-fe-conversation: global developer-mode raw-JSON switch (not added)	wiki-local/frontend/chat/developer-mode-raw-json-switch.md
180e2831e117e7aa	handfish	t8-fe-conversation: mock-source additions for the 4 missing kinds	wiki-local/frontend/chat/mock-source-showcase-scenario.md
5a4cb7ace8c18a04	handfish	t8-fe-conversation: where the new CSS for mention chips, the "응답 필요" badge, and the thread-side answer widget lives	wiki-local/frontend/chat/chat-css-ownership-blocks.md

Two rows in this set are near-general and were still excluded: 0f78e800933d2aaa (UnixListener::bind rejected because the fixture path exceeds macOS's 104-byte sun_path) is already covered by the bundled platforms-filesystems-unix-domain-socket-path-length; 1a92fa259582f25b (decline an optional settings surface the brief leaves open) is a product-scope call whose reasoning is the brief's, not a reusable trigger.