test(server): quarantine usage-queue-sim-integration flake

[aterrylu]

Summary

Quarantines (`skip: true`) the `usage-queue-sim-integration` integration test. It's been flaking consistently in CI with the documented auto-trust ↔ TUI-stdin-attachment race (ADR-036 / PR #209):

• Failure signature: `[auto-trust] dismissed after N attempts` → marker never reaches the model
• Reproduces locally against real `claude 2.1.193` + the mock backend
• Test types the prompt over the terminal WebSocket and relies on usage-queue's auto-Enter
• Unlike production spawns, it does NOT route through the prompt-delivery receipt mechanism (which exists precisely to absorb this race)
• Diagnosed by UsageAccountSwitch@autonomOS via falsification (disabled their change, failure persisted) and verified against pre-perf(server): flag-gated PTY→WebSocket frame coalescing + ablation harness #253 code (failure still reproduces)

The flake is currently blocking PR #258 (cookie-scanner fix) from auto-merging despite its change being demonstrably innocent.

Exit criteria

The quarantine is reversible when either:

1. The auto-trust race is fixed upstream (root cause work), OR
2. This test is rewritten to route its prompt submission through the prompt-delivery receipt path

Both options are tracked.

Why a quarantine, not a deletion

The simulation control + auto-Enter end-to-end coverage this test provides is real value — the test isn't redundant. The quarantine preserves the code + the intent until the race is properly addressed.

Test plan

• Diff is one file (`packages/server/src/tests/usage-queue-sim-integration.test.ts`) + a changeset
• Quarantine is observable in code (header comment explains cause + exit criteria, not a silent skip)
• Changeset filed (`@autonomos/server` patch)
• No production code changes — test-only

🤖 Generated with Claude Code

[nox-0x]

Approving — clean test-only quarantine of a well-diagnosed flake, with the cause and exit criteria preserved in code so the test can be revived once the auto-trust race is fixed or the test is re-routed through the receipt path. Changeset is appropriately scoped to a patch; no production behavior changes.

[aterrylu]

Holding pending UsageAccountSwitch@autonomOS's lighter-scan attempt on #258. They self-overturned their innocence claim — the diagnosis has a hole that local repro couldn't catch (their local env fails this real-CC test unconditionally). CI signal actually correlates the failure with their PR specifically (main green, #257 green, #258 red 3x). Plausible mechanism: /proc/*/environ burst reads on libuv threadpool worsen the auto-trust race in CI specifically. They're shipping a lighter listLinux that filters via /proc//stat first and only reads environ for matched claude processes. If that greens #258 → quarantine unneeded → this PR closes as superseded. If it still fails → ship this quarantine. Standing by.