[choiyounggi]

Knowledge flush — 2 insight(s) (headless auto-flush run 20261030-153306-64550)

35 queue rows were claimed this run: 2 ingested, 31 dropped as project-specific
plan-gap rows, 2 released because they plausibly overlap open PRs. Nothing was folded
into another branch.

The 31 plan-gap rows are lines 278–308 of ~/.dev-loop/queue/plan-gaps.jsonl. This
headless session cannot list the queue directory, so the rows were located by
filename: claim order follows sorted filenames and file order, the first and last six
claimed ids match lines 278 and 303–308, and the 31 ids fill that range exactly.

Verified best-practice

1. Freshness check of an uncommitted generated file (row 9950a83a55ea53e7).
   Claim: before a generated file's first commit, compare a saved copy with the
   regenerated file (cmp); git diff --exit-code after git add -N is always red.
   • https://git-scm.com/docs/git-add — -N: "An entry for the path is placed in
     the index with no content."
   • https://git-scm.com/docs/git-diff — the plain form shows changes "relative to the
     index"; --exit-code exits 1 on differences.
   • https://code.googlesource.com/git/+/HEAD/t/t2203-add-intent.sh — git's own
     tests expect an intent-to-add path to diff as a new file with all lines added.
   • Session evidence: the gate exited 1 with new file mode 100644 and 444 added
     lines on a byte-identical regeneration; the cmp check exited 0.
   • Confidence: verified (official docs plus a field reproduction). No command
     was re-run in this headless session, which has no shell for git.
2. Test module written during a background baseline run (row bc0e0b2a5d0c2dca).
   Claim: when a full-suite baseline runs in the background, wait for it or run it in
   a separate copy before creating test modules in the same tree; a test that
   re-discovers the tree at runtime imports the half-written file.
   • https://docs.python.org/3/library/unittest.html#unittest.TestLoader.discover —
     "If importing a module fails … this will be recorded as a single error and
     discovery will continue."
   • Session evidence: 2 test_readme_currency failures citing
     Failed to import test module: tests.test_mcp_exec_tools (the file being
     written); with the module complete, the full run reached 5959 OK.
   • Confidence: the mechanism is doc-backed; the directive is field-tested. It is
     added as one edge row on a verified page, and its source line marks it as a
     field incident.

Existing-layer check

Pages read: qa-document-verification-generated-reference-drift-gates, platforms-tools-staging-tracked-files-under-an-ignored-directory, testing-data-test-data-and-isolation

Also read the domain indexes wiki/testing/index.md, wiki/qa/index.md and
wiki/platforms/index.md, and ran wiki_search for each candidate's trigger.

• Row 1: no page covers intent-to-add or untracked files in a freshness gate.
  generated-reference-drift-gates owns --check gates for closed-vocabulary
  references (a narrower trigger), and staging-tracked-files-under-an-ignored-directory
  owns git add exit codes. → new page platforms/tools/freshness-check-of-an-uncommitted-generated-file.md,
  related-linked to both, to platforms-toolchains-regeneration-silently-drops-hand-edited-state,
  and to testing-quality-checks-that-cannot-pass.
• Row 2: no page covers a concurrent edit breaking a runtime-discovery test.
  testing-data-test-data-and-isolation owns "suite behaves differently because of
  its environment". → merged as one Edge-cases row plus two source lines, a
  related link to qa-process-scope-purity-checks (same root cause: a suite reading
  the live tree), and an extended load-when line in the testing index.
• No conflicts with existing directives.

Open-PR check

Open knowledge/* heads: #318, #317, #315, #310, #309, #308, #307, #306, #305, #304,
#301, #300, #299, #298, #293. This headless run read their changed-file lists
(gh pr list --json files), not their diffs.

Row	Overlapping PR	Verdict
9950a83a55ea53e7 git add -N freshness check	none — #293 (diff-algorithm line counts) and #306 (pip --require-hashes known-bad test) touch other mechanisms and pages	new
bc0e0b2a5d0c2dca test file written mid-baseline	none — #317 touches testing/data/artifact-leakage-from-a-suite.md (leftover files), a different trigger	new
fc76dc4dfa08ecde docker schema validator: skip vs REQUIRE fail, known-bad document	plausibly #300 (edits testing/quality/tests-that-cannot-fail.md and checks-that-cannot-pass.md)	released (pending duplicate)
122dd81477d4a447 absence-in-source assertion fails on its own explanatory comment	plausibly #317 ("self-matching gate CHECKs", edits checks-that-cannot-pass.md)	released (pending duplicate)

The released rows are left pending, not retired, so the manual flush can diff them
against those PRs.

Routing decision

• Row 1 → platforms/tools/freshness-check-of-an-uncommitted-generated-file.md (new
  page). It is a git-behaviour edge case that applies to any repository, which is
  the scope of the tools category; no new category needed.

• Row 2 → testing/data/test-data-and-isolation.md, Edge cases (merge).

• 31 plan-gap rows → not ingested (see below).

Local-layer candidates

31 rows from plan-gaps.jsonl (repo shag = linkly, task plans of orchestration
run oi1029; the 7 rows read in full come from t1, t5 and t9), ids 1d01cb971e07f6cb … 9bf49772c47a2672. Each records
one plan decision (Planning tN: deciding … (no owning wiki page)) that names
linkly's own files, flags and plan ids (cmd_token, docs/backends.md,
benchmarks/token/, plans/t9/design.md D23), so it would be wrong in another
codebase. Target, if anyone wants them kept: wiki-local/<domain>/<category>/<slug>.md
inside linkly — run wiki-ingest inside that project. Each decision is already
recorded in that run's plans/tN/design.md, so they were retired as
dropped: project-specific plan-gap (local layer of its repo), the same outcome
earlier flushes gave such rows.