test: pin the fleet hook stdin contract with a cross-plugin corpus

[gtapps]

Summary

Every fleet hook keeps its own sealed copy of stdin handling — fleet plugins can't import core at runtime, so the four invariants core states in scripts/lib/hook-input.ts (drain fully past the cap, size cap, fail direction, decision shape) were re-decided per file with nothing pinning them. Five distinct stdin idioms and four decision idioms ship across the fleet today.

This adds a repo-root corpus test that feeds all 11 gate hooks the same adversarial payloads on real stdin and asserts each hook's declared fail direction. Writing it immediately surfaced two live defects in claude-code-dev-hermit, both fixed here.

Changes

The corpus (tests/cross-plugin/hook-stdin-contract.test.ts)

• One spec row per gate declaring its fail direction. HA's mcp-safety-gate is the fleet's only fail-closed gate (exit 2); the other ten fail open.
• Cases: empty stdin, non-JSON garbage, non-object JSON ([1,2] / null / "str"), a well-formed benign payload, and a generated 2MB unparseable payload (generated, not committed — HA's gate-corpus precedent).
• Drain is asserted via the writer's exit code, piping through bash and reading PIPESTATUS. This matters: a hook that exits mid-stream and one that drains cleanly both report 0 themselves, so the hook's own exit code cannot detect the bug — only the writer's SIGPIPE (141) can.
• An auto-discovery test walks plugins/*/hooks/hooks.json and checks both directions: a stale/renamed spec fails, and a newly registered PreToolUse gate missing from the corpus fails too.

Two defects it caught (claude-code-dev-hermit)

• git-push-guard, worktree-boundary-guard, record-test-result all called process.exit(0) inside the read loop on oversize input, abandoning the pipe half-read — the exact failure core's header exists to prevent. worktree-boundary-guard even carried a comment claiming it drained to completion. They now stop buffering but keep consuming to EOF.
• record-test-result crashed with exit 1 on a null payload: its property access sat outside the parse try/catch, unlike its two siblings. Claude Code treats exit 1 as non-blocking, so this was an uncaught crash masquerading as fail-open.

Wiring

• test-cross-plugin.yml path filter now includes plugins/*/hooks/** and the three dev guard scripts — none of them triggered this workflow before.
• scripts/test-all.sh now runs the cross-plugin suite, which it never did. Deliberately serial after the parallel phase: these spawn a subprocess per corpus case, and the parallel phase is already saturated enough that HA's CPU-bound gate-corpus tests sit near their 5s per-test timeout.

Deliberately not done

• The envelope bytes stay out of scope. HA's two hookSpecificOutput emitters are duplicated on purpose and already byte-pinned by gate-corpus.test.ts against the retired Python gate. Re-asserting them here would duplicate a deliberate pin.
• No shared drainStdin() helper. A review pass proposed extracting the three now-identical loops into dev-hermit's scripts/lib/. Rejected: the natural helper hides process.exit(0) inside something that reads as "return a string", hardcoding fail-open into the helper and removing each caller's declared fail direction — the exact property this PR makes explicit. Core's own readHookInput returns an OVERSIZE sentinel precisely so the caller decides; doing it that way is a redesign of three hooks.
• No caps added to HA/forge/feed. readFileSync(0) and Bun.stdin.text() already read to EOF, so they satisfy the drain invariant; the corpus records cap: none rather than forcing a change.

Test plan

• bun test tests/cross-plugin/ → 141 pass, 0 fail.
• bunx tsc --noEmit → exit 0.
• bash plugins/claude-code-dev-hermit/tests/run-all.sh → exit 0.
• Mutation check: restoring the mid-loop process.exit(0) in git-push-guard.ts fails exactly one test, naming the hook and the oversize case, on the drain assertion (expected 0, received 141) while the hook's own exit code stays 0 — confirming the check is not vacuous.

Known unrelated failure

bash scripts/test-all.sh exits 1 on HA's two gate-corpus oversize tests (~6.1–6.6s vs a 5s per-test timeout). This is pre-existing and not caused by this branch — verified by stashing every change on this branch and re-running the untouched runner, which fails identically. HA's suite passes on its own (32s). Left alone as out of scope.