[prodelt]

Summary

The first audit of this harness run on Linux rather than Windows. On Linux and macOS — the only server platform, the one the harness is declared a core for — none of the three links in its own promise held, and none of the three reported a failure. Spec 024 is closed: ten defects, all reproduced, each fix shipping the regression that was missing.

	before	after
node tools/elt-oracle-runner.js --full	110/112, exit 1	118/118, exit 0
node tools/elt.js oracle	exit 1 (0s), no reason printed	full run, real exit code
elt commit with the shipped hook enabled	impossible by construction	passes
a suite run inside the gated project	wrote 5 fabricated ctx7 proofs	hermetic: tree byte-identical
a hung judge in background verify	exit 0, no run-log, no queue row	non-zero, terminal record written
a refusal read by an unattended agent	exit 4 for four different causes	a stable reason slug per cause
.harness/harness.json	8 of 33 fields validated	every field typed; typos rejected

Why it could not work outside Windows

• T001 (D29) — the platform is chosen, not assumed. The shipped config carried shell: "powershell"; sh() spawned it unchecked, and on ENOENT spawnSync returns status: null, which became an ordinary red oracle. The shell dispatcher existed in five copies, so a fix in one never reached the others. Now one module (tools/shell-run.js), default from process.platform, a missing interpreter names itself and the field to edit, and shell enters config validation.
• T002 (D33) — a hung judge fails closed. timer.unref() in withJudgeTimeout defeated the guard's entire purpose: an unref'd timer does not hold the event loop, so with a judge that never returns Node drained the loop and exited before the timeout fired — try/finally never ran and the process returned zero. The same line hangs the guard's own regression and cancelled 13 tests after it. The benchmark grader got a preflight: a missing pytest is no longer indistinguishable from a wrong solution.
• T005 (D32) — the gate lock actually hangs. .githooks/pre-commit was committed at mode 100644; git prints hint: ... not set as executable and commits anyway.

Why "green" did not mean "checked"

• T003 (D30) — the tree proof is computed from content, not from git's output format. treeHash() hashed git status --porcelain lines verbatim, status column included. That column changes on staging (" M f" → "M f", "?? g" → "A g"), and elt commit runs git add -A exactly between minting the proof and re-checking it in the hook — so it could never match. Second half of the same root: .trim() on the whole stream ate the leading space of the first line, and line.slice(3) then truncated the path.
• T004 (D31) — the oracle cache stopped being evidence. computeEntry was called without scanDirs while TEST_ROOTS has three entries, so editing bin/l0.js never moved the key of bin/l0.test.js — sabotaging the L0 entry point printed 65/65 passed in 2.4s. The cache also lived in a gitignored .harness/, invisible to treeHash. Now the roots are shared, RUNNER_VERSION covers the selection rules, the cache moved to .git/elt/, and the proof carries ran/cached — ran: 0 on a non-empty selection is refused as an absence of checking.
• T006 (D34) — the suite stopped issuing permits to itself. The live ctx7 ledger was 75 rows, 100% fabricated, refreshed on every oracle run so the 30-day freshness window never expired, and substring matching let them cover ai, mcp, agents, playwright, cli and vercel.
• T010 (D38) — three checks that could not fire. Two filtered the review queue on a resolved field nothing writes (rows close with closedAt) — in graph-state a settled row could block elt cutover forever. The --push guard read commitHash from certificates that carry commit; a naive rename would have flipped it to always-block, since the certificate stores a short sha and the check used the full one.

Why an outside developer, or a server agent, could not use it

• T007 (D35) — package.json now exists, with engines, npm test, and the version as a fifth source in version-check (the root of D28).
• T009 (D37) — a refusal is now readable by a machine. judge-invoke.js dropped r.reason, so reasons arrived empty on failure; judgeLog reached elt.js and went only to the run-log. The real cause lived in a file nothing printed — on a server that is --dangerously-skip-permissions cannot be used with root/sudo privileges. And exit code 4 covered 44 of 55 refusals. Renumbering would break drivers that branch on === 4, so the discriminator became a slug: die(msg, code, reason) plus {ok, code, reason, message} under --json or ELT_JSON=1. --json now covers commit, gate and judge run. Human output is unchanged byte for byte, and a regression asserts that.
• T008 (D36) — the config is a contract. Eight of thirty-three fields were validated; the rest did not reject a typo, they changed the gate: redProof: "OFF" kept the circuit on, specApproval: "no" kept the approval gate on, batch: "three" silently became 3. Types are errors now; unknown keys are warnings until the next minor (SCHEMA_VERSION), because retired fields sit in existing configs. The schema was written from the code, which paid off immediately — its first version rejected working configs on branchPolicy, where an invented current disagreed with the real feature | none.

Coverage 112 → 118. Full evidence in specs/024-harness-scale-readiness/audit.md (E1–E22); D29–D38 closed and D39–D41 plus the D36 remainder opened in docs/DEFECTS.md.

Test plan

Measured on Linux, Node 22.22.2, clean tree. CI ran the same suite green on windows-latest and ubuntu-latest for every commit on this branch.

node tools/elt-oracle-runner.js --full   # 118/118 passed in 49.3s — exit 0
node tools/gen-agents-md.js --check      # копии совпадают с CLAUDE.md — exit 0
node bin/doctor.js                       # exit 0
node tools/version-check.js              # ok — 5.0.1 (five sources)

Hermeticity verified separately: git status --porcelain -uall hashed before and after a full run is byte-identical, and .harness/ctx7-proof.jsonl is no longer created.

The branch's smallest regressions:

• tools/elt-tree-hash.test.js — discriminating, verified by reverting the fix: 3 of its 5 assertions go red on the pre-024 elt.js. The other two guard against over-fixing — a real content change after the proof must still be refused.

• tools/shell-run.test.js — a missing interpreter is named rather than reported as a red oracle; the shipped config carries no hard powershell; no second dispatcher copy may reappear.

• tools/supply-exec-bit.test.js — every file under .githooks/ is executable in git's index (checked there, not on the filesystem, so it holds on Windows too).

• tools/elt-oracle-cache.test.js — editing a source under bin/ or benchmarks/ moves its test's cache key; SCAN_DIRS and TEST_ROOTS may not diverge again.

• tools/elt-commit-proof.test.js — a proof with ran: 0 is refused on both the plain and the trusted gate path; an old-schema proof is refused by name.

• tools/context7-cli.test.js — an injected runner mints no proof; coverage matches package names exactly, and the free-text query no longer counts at all.

• tools/elt-machine-readable.test.js — two refusals sharing exit 4 are told apart by slug; ELT_JSON=1 works without the flag; without --json stdout stays empty and stderr is unchanged; a dead judge prints the provider log.

• tools/elt-dead-checks.test.js — each of the three revived checks is asserted to fire and not to fire where it shouldn't.

• tools/harness-schema.test.js — the exact config that used to validate is now rejected; retired fields are named as retired; the repo's own config passes with zero warnings.

• tools/review-runtime.test.js — a sixth lens no longer kills the judge, while a missing declared lens still does.

• full suite exits 0

• instruction drift check exits 0

• the smallest regression for this branch exists (or the change is trivial)

🤖 Generated with Claude Code

https://claude.ai/code/session_01ADgXkZxYt9KBYyjnPGCQeF