knowledge: 6 insights — input-manifest freshness with skipped inputs, uv stores under a relocated HOME (2 new pages), backtick span in a double-quoted grep pattern (+1 merge, 1 fold onto #183, 2 drops)

[choiyounggi]

Knowledge flush — 6 insight(s)

Claimed queue ids: bf675f09e7973bbc, 6a0cdc6a27637c93, ad50053952f28169,
354053be0de8d411, aca4c8fce4cabd8a, 66894645015082e9.
All six were handled: 2 new pages, 1 amended page, 1 folded onto open PR #183, 2 dropped.
Branch knowledge/choiyounggi-20260917-160110, flush run 20260917-160035-87121.

Verified best-practice

1. bf675f09e7973bbc — a freshness manifest must record skipped inputs too (→ new page, confidence: verified)

Claim: when a derived index judges freshness by comparing a stored input-hash manifest
with the current input set, and the builder may skip an unprocessable input, the skipped
input's hash still belongs in the manifest (it is excluded from the output table only).
A manifest filled from successes can never equal the current inputs, so status reads
stale straight after a successful build and every trigger rebuilds — with no error
logged, because each rebuild succeeds.

• A persistently failing file makes every incremental update a full rebuild (identity-version loop) tirth8205/code-review-graph#944 — fetched this session via
  gh api. An independent project hit the same loop in its version-marker form: the
  builder refuses to record its identity version when any file failed to parse, so
  "every incremental update becomes a full rebuild, indefinitely … it is silent apart
  from an INFO log". Its proposed direction matches the directive: "Record the version
  and track the specific failed files, so only those are retried".
• https://aws.amazon.com/builders-library/caching-challenges-and-strategies/ — fetched
  this session. Negative caching ("cache the error response … using a different TTL than
  positive cache entries"; "Don't cause or amplify an outage by repeatedly asking for the
  same downstream resource and discarding the error responses") is the remote-call form
  of recording a failed input. Used for the principle only, not for the manifest shape.
• Own reproduction this session (Python 3, two-file fixture, three simulated sessions
  that rebuild when stale): success-only manifest → builds over 3 sessions: 3 | final status: stale; all-inputs manifest → builds: 1 | final status: fresh; identical
  output table in both runs.
• Session evidence corroborated read-only in the originating worktree: the index builder
  carries the fix (docstring "A page that cannot be parsed still gets its sha recorded")
  and tests/wiki-index.bats has the case "an unparseable page still lets the index
  settle to fresh" with a repair-the-page negative control. That work is uncommitted in
  its own worktree, so I read it and did not run it there.

2. 6a0cdc6a27637c93 — uv under a relocated HOME needs both UV_CACHE_DIR and UV_PYTHON_INSTALL_DIR (→ new page, confidence: verified)

Claim: uv resolves its package cache and its managed interpreters from two separate
HOME/XDG-derived locations. Carrying only the cache across a HOME swap makes uv fall back
to a different Python, so --offline fails on packages that are cached; a suite that
turns that failure into skip then passes without executing the cases.

• https://docs.astral.sh/uv/reference/storage/ — fetched this session. Cache:
  $XDG_CACHE_HOME/uv or $HOME/.cache/uv. "By default, Python versions managed by uv
  are stored in a python/ subdirectory of the persistent data directory, e.g.,
  ~/.local/share/uv/python" — overridden by UV_PYTHON_INSTALL_DIR. Tools: tools/
  under the same data dir, overridden by UV_TOOL_DIR.
• https://docs.astral.sh/uv/concepts/cache/ — fetched this session. Precedence:
  --no-cache temp dir, then "--cache-dir, UV_CACHE_DIR, or tool.uv.cache-dir",
  then the system default. The page has no sentence on --offline, so the page body
  makes no documentation claim about --offline beyond the field measurement.
• Own reproduction this session (uv 0.11.5, macOS arm64): HOME=<scratch> moved both
  uv cache dir and uv python dir under the scratch dir and removed the managed
  cpython-3.14.4 / cpython-3.12.13 from uv python list --only-installed (only
  Homebrew and python.org interpreters remained); exporting both variables with the
  pre-swap values restored both locations. XDG_CACHE_HOME / XDG_DATA_HOME alone moved
  the cache and the python/tools dirs respectively.
• Not reproduced here: the exact "not found in the cache" failure (it needs the
  originating project's cache contents). It is recorded on the page as a field
  measurement, not as a documented behavior. A Windows inference I first drafted
  (which variable relocates %LOCALAPPDATA%\uv\cache) was removed as unverifiable on
  this machine; the row now states only the documented default.

3. ad50053952f28169 — a backtick span inside a double-quoted grep -F pattern is executed, not matched (→ merged into an existing verified page)

• https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html — fetched
  this session. 2.2.3: "The backquote shall retain its special meaning introducing the
  other form of command substitution"; 2.2.2: "A single-quote cannot occur within
  single-quotes."
• Own reproduction this session under sh, bash, zsh and dash: the double-quoted
  raw-backtick pattern printed wiki-local/: No such file or directory and grep -qF
  returned 1 with the text present; printf '[%s]' showed grep received [own layer].
  Single-quoted, \`-escaped inside double quotes, and grep -F -f file each
  returned 0.
• Correction to the candidate's evidence line: as harvested it shows the pattern with
  \` escapes, which the reproduction proves would have matched. The originating
  gate file (read this session) now holds the single-quoted form, with the pattern
  shortened to avoid the apostrophe in "project's" — that apostrophe case is why the
  merge adds a second row (-f file / shorten the span) rather than only "single-quote it".

4. 354053be0de8d411 — per-site mutation of a source-text wiring test (→ dropped, already covered; claim itself is sound)

The directive (enumerate the protected sites, delete each one, require red) is already
the verified content of testing-quality-source-text-wiring-assertions steps 1 and 5,
sourced to Stryker and PIT. No new verification was needed.

5. aca4c8fce4cabd8a — coordinator, not the design author, invokes the plan reviewer (→ folded onto #183, field-tested at most)

• https://code.claude.com/docs/en/sub-agents — fetched this session; all three quoted
  platform facts in the candidate are present verbatim ("Claude uses the SendMessage
  tool with the agent's ID or name as the to field to resume it"; "up to three layers
  below the main conversation"; "omit Agent from its tools list or add it to
  disallowedTools").
• The handshake itself is a plan decision with no measured outcome, so it is not
  evidence for a verified claim; stated as such in the fold comment.

6. 66894645015082e9 — README wording for the task-planner agent (→ dropped, unverifiable as a best practice)

A repository-specific wording decision (which README line gains which sentence). It has
no trigger outside this repo and no directive a future task elsewhere could apply.

Existing-layer check

Pages read: backend-common-caching-invalidation-and-stampede, platforms-tools-version-keyed-artifact-cache, testing-quality-source-text-wiring-assertions, platforms-shells-escapes-in-shell-string-literals, platforms-shells-command-text-inspected-before-execution, platforms-shells-portable-shell-scripts, testing-data-test-data-and-isolation

Also consulted: INDEX.md, wiki/backend/index.md, wiki/platforms/index.md; whole-wiki
greps for manifest, UV_CACHE_DIR|UV_PYTHON_INSTALL_DIR, uv run|uv sync,
backtick|backquote|command substitution, mutat, source-text|wiring test,
self-grad|fresh-context (all hits listed, none truncated); grep-only look at
mutation-harness-file-custody and tests-that-cannot-fail for restore-verification and
skip coverage.

• knowledge: streaming large result sets (server-side cursor blocked → fetchmany + write-only) #1 manifest: invalidation-and-stampede covers read caches (TTL, delete-on-write,
  stampede, negative caching) and version-keyed-artifact-cache covers a version-string
  key that never changes. Neither covers a per-input manifest with skipped inputs — new
  trigger → new page. Linked both ways with invalidation-and-stampede (rule 6,
  negative caching, is the remote-call analogue); one-way to version-keyed-artifact-cache.
• knowledge: kubelet resource metrics on embedded k8s; host-to-pod memory limit sizing #2 uv: no page mentions UV_CACHE_DIR or UV_PYTHON_INSTALL_DIR (0 hits).
  test-data-and-isolation has the row that tells you to point HOME at a scratch dir —
  this page is the consequence for a tool that locates its state through HOME. New
  trigger → new page in platforms/toolchains. No conflict: the new page keeps HOME
  relocated and overrides uv's stores, consistent with that row.
• knowledge: PID-1 tee log capture pattern, BSD/GNU date %3N feature detection #3 backticks: portable-shell-scripts step 7 already states the rule for message
  text, and escapes-in-shell-string-literals owns pattern literals. Same trigger family,
  same directive (single-quote) → merged into escapes-in-shell-string-literals: one
  "When this applies" sentence, two edge-case rows, one Instead of row, two source
  bullets, last_verified bumped. No conflicting directive.
• knowledge: kubelet resource-metrics, PID1 log flush, date %3N detection #4 per-site mutation: duplicate of source-text-wiring-assertions steps 1 and 5 and
  its Instead of row "Ship the anchored guard because the suite is green". The only
  nuance (the sites were documentation sentences) is covered by that page's doc-gate edge
  row. The page is also at its body-line budget. → no edit.
• Related-links deliberately one-way: test-data-and-isolation, version-management
  and environment-resync-removes-undeclared-packages are each modified by open PRs
  (knowledge: 12 insights — deny rules under bypass, merged-tree gate, worker verify command, Kotlin daemon heap, extracted-method this (+7 merges) #179, knowledge: 12 insights — dropzone copy vs drop handlers, destination-in mask chaining, media-query inset reset, spatial clamp, env restore vs pop, synthetic-corpus floor, plan-claim recompute, multi-name command -v, alert suppression key, sibling validators (+2 folds) #181, knowledge: 21 insights — fork-PR secret gates, review-bot exit codes, dict-subclass copy loss, iterative DFS, citation sweep, plugin-prose identity leak, xcodegen regen loss, narration ordering tests, autouse shadowing, stub contracts, real-CLI spot checks (+8 merges, 1 fold onto #183) #188, knowledge: 3 insights — greedy-regex adjacent tokens in extractor fixtures (1 new page), subagent model-scoped 429 → per-invocation model override (+1 merge, 2 same-batch duplicates folded, Sonnet-limit drift fixed) #191), and their related: lines are the usual conflict point, so
  the uv page links to them without a back-link in this PR. escapes-in-shell-string-literals
  is touched by knowledge: 6 insights — gate EVIDENCE exit-code class (unset CLAUDE_PLUGIN_ROOT), jq . rebinding in predicates, worker-reported plan contradiction, benchmark speedup attribution, sequential-dispatch test under a thread pool, lock-protection race-window test (6 new pages, 0 merges, 0 drops) #189 on its related: line only; this PR leaves that line unchanged.
• Checks run in the checkout: node scripts/wiki-structure-checks.js wiki →
  pages: 280, indexes: 13, findings: 0; node scripts/wiki-lint-prohibitions.js wiki →
  directives: 75, violations: 0 (75 before and after, so the bats pin is unchanged);
  bats tests/wiki-lint-prohibitions.bats tests/wiki-structure-checks.bats tests/wiki-lint-model-era.bats tests/wiki-lint-score.bats → 1..38, 38 ok, 0 not ok.
  New page bodies are 68 and 71 lines; the amended page is 67.

Independent review before commit (8 files changed). A fresh-context reviewer with an
adversarial brief re-fetched every cited URL, re-ran the uv and backtick reproductions
under sh/bash/zsh/dash, re-ran the linters and the 38 bats cases, and checked this
report's PR-number claims against the fetched heads. Verdict PASS-WITH-FIXES, three
findings, all applied: (1) the manifest page's "retry only when the hash changes" was
overstated — a builder upgrade that can now parse a skipped file leaves its hash
unchanged, so an edge row now keys retry on (input hash, builder version); (2) and (3)
the two field-measurement bullets read as if executed in this flush — both are now
labelled "Reported field measurement … not re-executed / not independently reproduced",
and uv step 2 separates the reproduced fallback from the reported --offline failure.
The reviewer left the working tree byte-identical (diff compared against a pre-review
snapshot).

Open-PR check

Listed with gh pr list --repo choiyounggi/dev-loop --state open --search "head:knowledge/":
#179, #180, #181, #182, #183, #185, #186, #187, #188, #189, #190, #191, #205, #207, #208
(15 heads). Each head was fetched and its origin/main...head -- wiki/ diff grepped per
candidate; every hit file was listed.

Candidate	Overlapping open head	Verdict
bf675f09e7973bbc manifest	none — manifest/fresh hits in #183, #185, #186 are unrelated (artifact cache version bump, graph mtime gate, key regeneration)	new
6a0cdc6a27637c93 uv	none — sole uv run hit is a verify-command table cell in #179	new
ad50053952f28169 backticks	none — 0 added lines match `backtick	backquotein any head; #189 edits only the target page'srelated:` line
354053be0de8d411 per-site mutation	none in flight; duplicate of the merged layer	drop
aca4c8fce4cabd8a reviewer ownership	#183 adds qa/process/fresh-context-code-review, whose trigger is "designing which session or subagent an automated review stage dispatches to"	fold — unique addition (nested-spawn edge row, coordinator-owned handshake) noted on #183: #183 (comment)
66894645015082e9 README wording	none	drop (project-specific)

No sibling duplicate PR is opened by this flush.

Routing decision

Candidate	Target	Decision
bf675f09e7973bbc	backend/common/caching/input-manifest-freshness-with-skipped-inputs	New page in the existing caching category; wiki/backend/index.md row added
6a0cdc6a27637c93	platforms/toolchains/uv-state-directories-under-a-relocated-home	New page in the existing toolchains category. The candidate's hint was testing, but the knowledge is where a tool keeps per-user state, which is the platforms domain; the testing-side rule (relocate HOME) already lives in test-data-and-isolation and is linked. wiki/platforms/index.md row added
ad50053952f28169	platforms/shells/escapes-in-shell-string-literals	Merged; that page's index "load when" line extended with the backtick symptom
354053be0de8d411	—	Dropped as covered by testing/quality/source-text-wiring-assertions
aca4c8fce4cabd8a	qa/process/fresh-context-code-review (on #183)	Folded by PR comment; not ingested here
66894645015082e9	—	Dropped as project-specific

No new category was created.

[choiyounggi]

Merged into main via the consolidated knowledge backlog: this PR's content landed in #221 (part of the 6-part stack #216–#221, reviewed as a whole in #215). Conflicts with sibling knowledge PRs were resolved by union, and any review fixes are listed in #215 and the part PR. Closing as absorbed.