Knowledge flush — 7 insight(s): 6 ingested (1 new page, 5 merges), 1 dropped

Verified best-practice

1. Policy gates must parse all three shell quoting forms and expand only resolvable prefixes (d3ef0214, from dev-loop)
Claim: a PreToolUse/policy gate extracting an argument from raw command text must accept bare, single-quoted, and double-quoted forms and expand ~/$HOME/${HOME} itself; a bare-token regex denies well-formed commands with a misleading "argument missing" error.
Sources checked: https://code.claude.com/docs/en/hooks (hook receives the unexecuted tool_input.command string), https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html (2.2 Quoting defines exactly escape/single/double-quote forms; expansion happens only when the shell processes the line).
Verification: re-ran the repo's own regression suite this session — bats tests/pre-flush-pr-gate.bats tests 12 (double-quoted --body-file recognized) and 13 ($HOME path expanded) pass on current main. Confidence: verified.

2. Invoke helper scripts via sh "$SCRIPT" so test stubs need no exec bit (EDR-safe seam) (cc8801d1, from dev-loop t1)
Claim: designing production code to call external scripts through the interpreter lets tests inject stubs as plain read-only files — no chmod +x, which EDR agents (SentinelOne) flag on temp/test paths — and survives mode-stripping distribution paths (plugin cache).
Sources checked: https://pubs.opengroup.org/onlinepubs/9699919799/utilities/sh.html — command_file "need not be executable" (stated for both slash and no-slash pathnames).
Verification: local reproduction 2026-08-06 — a mode-644 script ran via sh file (exit 0) and failed direct exec with permission denied (exit 126). Original field evidence: 8 stubs injected without chmod, 331/331 bats green. Confidence: verified (mechanism doc-backed + reproduced; EDR angle annotated as field practice).

3. Gate warnings by captured stderr, not exit code; redirection order 2>&1 >/dev/null (e3b04b0e, from linkly)
Claim: warning-emitting tools exit 0, so a feedback hook must capture stderr (OUT=$(tool "$F" 2>&1 >/dev/null)) and, in a Claude Code PostToolUse hook, forward non-empty output via exit 2.
Sources checked: https://code.claude.com/docs/en/hooks — PostToolUse "Shows stderr to Claude; the tool already ran" on exit 2; exit-0 stderr "goes to the debug log only, never the transcript". https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html — 2.7 Redirection: "the order of evaluation is from beginning to end". (gnu.org bash manual returned HTTP 429 repeatedly; POSIX covers the same semantics and is cited instead.)
Verification: local reproduction 2026-08-06 — cc -Wall on unused-variable code: exit 0 with 157 bytes on stderr; correct order captured the 156-char diagnostic, reversed order captured 0 chars. Confidence: verified.

4. Attribute temp-artifact leaks by prefix histogram before fixing; then a static cleanup guard observed red first (a61dfd51, from linkly)
Claim: histogram surviving temp names by prefix and match against mkdtemp call sites — the distribution pinpoints the few offending sites; enforce the cleanup convention with an AST/lint guard proven red on pre-fix code.
Sources checked: https://docs.python.org/3/library/tempfile.html — "The user of mkdtemp() is responsible for deleting the temporary directory and its contents when done with it."
Verification: field measurement (998 leaked dirs = 686+306 under two prefixes, exactly the 2/6 call-site files without cleanup; post-fix suite temp delta 0, 72 MB → 3.3 MB). The histogram-attribution method itself has no external source. Confidence: field-tested (cleanup-responsibility claim doc-backed).

5. A pre-implementation test that cannot go red: record vacuous green, prove later by guard mutation (d245e299, from dev-loop t2)
Claim: when a usage-error test expects the exit code the unimplemented path already produces, red-first silently fails; mark it vacuously green and after implementation mutate the specific guard, require red, restore.
Sources checked: existing page sources apply directly — James Shore AoAD2 TDD ("predict how it will fail") and https://testing.googleblog.com/2021/04/mutation-testing.html (already cited on the two target pages; both live-verified in prior flushes).
Verification: session reproduction — sed mutation of the arg-count guard (-ge 2→-ge 1) flipped the test to not ok; restore flipped it back. Confidence: verified.

6. Client-side throttles must count token/auth issuance requests and stamp at actual send (58d7e79d, from auto-trading-bot/stock-trader)
Claim: auth refresh inside _headers()/interceptors bypasses a throttle layered above it; token POST + first API GET land in the same second, deterministically exceeding a 2-req/s cap only on cold-token days — presenting as intermittent failure.
Sources checked: searched for an official cross-provider statement that token-endpoint calls count toward rate limits; Auth0/Okta rate-limit docs confirm token endpoints are themselves rate-limited but state no general counting rule, so no external URL is cited for the directive itself.
Verification: provider-log evidence from the field incident (token POST :00.354 → issuance :00.495 → rate-limited API call :00.543 on both token-issuance days; cache-valid days clean). Confidence: field-tested — the merged row is annotated as such inside the otherwise-verified page, following the existing permissions-and-exec-bits mixed-provenance precedent.

7. dev-loop orchestrate worktree_escape fires on read-only cross-worktree access — budget escalation round-trips (28fd6dfe) — DROPPED (see Routing decision).

Existing-layer check

Pages read: INDEX.md; domain indexes for platforms, testing, backend, infrastructure; full pages command-text-inspected-before-execution, portable-shell-scripts (headings), permissions-and-exec-bits, tests-that-cannot-fail, checks-that-cannot-pass, test-data-and-isolation, what-to-mock (index line), timeouts-and-retries. Grepped the whole wiki for guardrail/escalation/orchestrate coverage (none).

• Insight 1 overlaps command-text-inspected-before-execution — that page covered the caller side only; merged the gate-author side as new step 6 + one Instead-of row + field context. No conflict: caller-side rows ("a gate that excludes quote characters cannot receive a quoted path") remain correct defensive guidance against naive gates.
• Insight 2 overlaps permissions-and-exec-bits, which already had "invoke through the interpreter" as a mode-stripping workaround; merged the EDR/test-seam design angle as one edge-case row + one Instead-of row + POSIX source. No conflict.
• Insight 3 had no owning page (checked portable-shell-scripts — portability scope, wrong home; command-text-… — PreToolUse command inspection, different mechanism). Created a new sibling page; related: linked both ways to command-text-inspected-before-execution and portable-shell-scripts.
• Insight 4 overlaps test-data-and-isolation ("Filesystem / temp files" row existed but only as per-test practice); merged the attribution/enforcement case as one edge-case row with an inline sanctioned hop to checks-that-cannot-pass (guard-red-first). Frontmatter related not cross-added — adjacency is one directive, not page-level.
• Insight 5 fit both tests-that-cannot-fail (mutation proof) and checks-that-cannot-pass (target-does-not-exist scope). Merged into checks-that-cannot-pass — its existing "test for behavior you are about to implement" row is the exact parent case — with an inline link to tests-that-cannot-fail for the mutation mechanics. The two pages already related:-link each other.
• Insight 6 had no rate-limiting page; timeouts-and-retries owns outbound-call discipline and already handles 429s — merged as one edge-case row + field-incident source line rather than creating a single-row client-rate-limiting page.

Routing decision

No new categories created; no contradictions flagged. Indexes updated: platforms (2 load-when lines + 1 new page row), testing (2 load-when lines), backend (1 load-when line). log.md appended.

Note: opened via the GitHub MCP connection because the local gh CLI token has expired (gh auth login needed); the required INGEST_REPORT.md ships in this PR at .dev-loop/INGEST_REPORT.md.