fix(memory-integrity): the budget constant was 585 bytes too generous; port the phantom-tool and fleet-size legs

[u00dxk2]

lib/memory-integrity.mjs had drifted behind the internal version it was extracted from. This closes the gap.

The bug

DEFAULT_INDEX_BUDGET_BYTES read Math.floor(24.4 * 1024) = 24,985. The loader's documented "24.4KB" is 24,400 decimal bytes; re-typing it as KiB gave the check 585 bytes of slack it was never granted — in the only direction that matters. An index between 24,401 and 24,985 bytes is silently truncated by the loader and passes this lint clean.

The existing test asserted DEFAULT_INDEX_BUDGET_BYTES + 1, which passes just as happily on the wrong constant. That is why it never caught it. The replacement pins the literal and both sides of the boundary.

Also ported from origin

• phantom-tool (opt-in via a new toolResolver option) — an index line that NAMES a tool or script which resolves to no file. The live instance: an index naming an "open-items projection" tool that never existed, because a compression pass kept the name and dropped the inline recipe. Unfollowable for eleven days, and every reader saw a capability the fleet didn't have. Extraction (extractToolReferences) is pure; existence is judged by a caller-supplied resolver, so the file stays fs-free. No resolver → the check doesn't run. Resolver returns null or throws → can't judge → no finding.
• classifyMemoryIndexSizes + MEMORY_SIZE_STATES — the fleet-wide read, which has a failure the single-dir lint does not: an agent whose index you could not measure. MISSING is fail-closed and a finding (the dir exists, so sessions run there with no index at all); a dir that never existed is declared, not swept, and not a finding; sweptCount counts only measured agents, so a run that measured nothing returns "nothing-swept" rather than "clean". Findings dominate — a one-agent fleet whose index is missing has zero measured agents, and calling that "nothing-swept" would bury the loudest instance of the exact failure the classifier exists to catch.
• LOADER_TRUNCATION_SENTENCE — one sentence, used by both budget messages, so every surface says the same thing about why the byte count matters: nothing errors, nothing warns, the tail is just gone.

Deliberately not ported

mungeProjectPath and stripWorktreeSuffix. Both encode Claude-Code-specific path shapes (~/.claude/projects name munging, .claude/worktrees suffixes), and both are about locating files — which this library does not do.

Checks

188 tests pass, 0 skipped. Five new: the pinned budget boundary, extractToolReferences dedup/URL/placeholder handling, the phantom-tool resolver contract across all five resolver outcomes, and the two classifyMemoryIndexSizes verdict cases.

Independent of #1 — branched from main.

🤖 Generated with Claude Code

[u00dxk2]

Codex adversarial round — 7 findings, 6 as written, 1 adapted

All seven reproduced against the first commit before anything was touched. Folded in as one commit; each test names the number the old code returned.

#	Finding	Verdict
1	README quickstart used statSync().size; on CRLF that's +1/line, so the documented integration over-reported at the boundary it had just pinned	as written — measure the read text, LF-normalized
2	scripts\foo.py → foo.py; file:bar.py → bar.py. The extractor invented tokens the index didn't contain	as written — \ is a separator; a token after : is skipped
3	bytes: -1 → OK with 24,401 bytes of headroom	adapted — non-negative safe integers only, else MISSING (fail-closed). Zero stays a measurement: an empty index is under the cap, and the JSDoc now says how that differs from the lint's indexRead
4	dirExists:false + real bytes discarded the measurement	as written — the measurement wins
5	Duplicate agents counted twice; empty name → clean row; null row → MISSING finding about a dir nothing is known about	as written — malformedCount, skipped, never measured, never a finding
6	TOOL_REF_RE quadratic: 30k hyphens = 2.5 s	as written — bounded segments + linear pre-check; 50k = 0 ms
7	README said the lint finds phantom tools; quickstart passed no resolver; an async resolver silently never fired	as written — row says opt-in + synchronous, quickstart shows one, a thenable throws TypeError

194 tests pass. The spawn EPERM Codex hit running npm test was its sandbox refusing node:test's child processes; --test-isolation=none gave the same 188 it was reviewing against.