refactor: isolate monitor receipts and expand fuzz coverage

[trvon]

Stack

#147 → #148 → #149 → #150. Each child targets the preceding branch; #147 targets master.

Scope

Stack 3/4. Base: fix/monitor-attachment-races. Isolates internal callback receipts and expands generic monitor coverage.

• MonitorManager.runCallbacks supplies a fresh deep snapshot to each consumer.
• Generated runtime and inactivity models check terminal tuples, callback eligibility, finished payloads, output bounds, batching, exit grace, cancellation, pruning, and shutdown.
• Native POSIX shell cases use owned gate barriers and a work file to distinguish output markers, physical exit, fixture-body execution, and authoritative monitor outcomes.
• Steering fixtures use memory storage rather than consume unrelated persisted-session recovery notices.

Evidence and classification

Internal consumer-isolation hardening: unchanged-source seed 130363 failed at run 6 and shrank to a terminal callback mutation followed by timeout stop. Replay path: 5:4:4:4:7:7:7:7:7:7:7. The reducer guard remained correct; a mutable callback argument bypassed it. The same minimum and nested/zero-forgery expectations pass after repair. Four passive controls remain green. The minimized replay executed 59 checks, not the requested 1,000.

Three 1,000-run generated campaigns pass: 4,779 lifecycle actions, 4,779 consumer actions, and 4,844 inactivity actions. These overlap; 14,402 is an action-execution count, not unique scenario coverage. Defaults: 100 runs; maximum: 1,000; up to 40 actions.

Native seed 155921: 64 executions, 24 raw exit-zero/no-gate-start witnesses, no fixture-body execution in those witnesses. Their physical outcomes include zero, 143, and SIGTERM. Monitor metadata stays stopped/manual or stopped/timeout, with no successful done event. This is producer-pattern compatibility, not attribution of the original shell producer or a monitor product defect. Default native runs: 8; maximum: 64.

The harness coupling was reproduced with a test-owned foreign event store. Isolation passes all five unchanged assertions and preserves primary/previous snapshot hashes. No real session files were deleted or product recovery behavior changed.

Limits

get() remains mutable internal access. This does not promise public API immutability or isolate reentrant get mutations. Controlled-child models exceed ordinary process ordering; native checks are POSIX-only and the local observations are macOS. Windows-native checks are unavailable, not passes. Cross-monitor behavior, clock skew, independent surviving descendants, original Pi UI/crash reachability, and original producer attribution remain unproved.

Fixture cleanup retains the primary trace and cleanup failures and attempts every cleanup job. Reports retain successful observations and the final shrunk failure trace, not every failed attempt.

Validation

Fresh serialized gate at this exact tip: 80 files / 1,261 tests; 11 property-suite files / 30 overlapping tests. Coverage: 90.66% statements, 85.21% branches, 96.44% functions, 95.28% lines.

All commands below exit zero. Lint has two established warnings. Package smoke checks 124 files and public types; audit reports zero vulnerabilities.

npm run lint
npm run typecheck
npm test
npm run test:property
npm run test:coverage
npm run build
npm run test:package
npm audit --audit-level=moderate
git diff --check

The first four-tip concurrent run also hit a backlog filter-supersession assertion. The unchanged index/lifetime/store focus passes 98 tests; the serialized full gate and coverage run pass. The transient is disclosed, not attributed to a fixed product defect.

Commits

• 83ca1bc — callback receipt snapshots and generated models.
• 6fce3cf — steering fixture storage isolation.
• 83d6f1d — native producer-pattern corpus.

Signed commits. Dependencies and product version unchanged. External evidence and .vitest/ are excluded.