Summary

Adds focused regression coverage for residual high-risk seams on main beyond open coverage PRs #298–#300. No intentional production behavior changes.

Risky behavior now covered

• SocketSendQueue failure paths — send === -1 queuing, non-serializable payload drops (serializationFailures), ws.send throws (sendFailures, no queue), and flush that drops a failing head then continues draining.
• Deferred tool-arg projection — partial tool_input_delta never publishes args during input-streaming; late deltas after terminal tools leave outbound/terminal args unchanged.
• 16ms micro-batch timer flush — rapid assistant deltas coalesce without a text_end boundary.
• Codex app-server client hang recovery — request timeout rejects without hung waiters; process exit rejects in-flight requests with pending method context.
• setModel while busy — mid-turn model switch emits busy and leaves the model unchanged.
• Windows sandbox setup write soft-fail — blocked destination .sandbox does not throw or claim an update.
• PI rate-limit abort gates — mid-attempt abort wins over retry; cancel-shaped messages skip retry; stopReason=aborted with a rate-limit message still retries; lifecycle-only parts remain retryable.
• Rate-limit classifier edges — response.statusCode, cause-over-lastError, depth-4 chain cap, and maxRetries NaN/Infinity/negative/fractional sanitization.
• Cold-start MCP close errors — close failures during sibling abort are logged and do not mask the original failure.
• Runtime attestation binds — incomplete or foreign-signature .verified.json fail closed on the verify fast path and are rewritten after a successful re-verify.

Test files added/updated

• test/server.backpressure.test.ts
• test/jsonrpc/projectors/assistant-segments.test.ts
• test/jsonrpc/projectors/tool-reuse.test.ts
• test/providers/codex-app-server-client.test.ts
• test/session/agentSession.model.test.ts
• test/platform/windowsSetupSync.test.ts
• test/runtime.pi-rate-limit-retry.test.ts
• test/agent.test.ts
• test/coworkRuntime.test.ts
• apps/desktop/test/research-readiness.test.tsx (desktop bridge CI stabilizer)
• apps/mobile/bun.lock (canary Bun frozen-lockfile stabilizer)

Why these tests materially reduce regression risk

These paths sit on shared concurrency, trust, and transport seams: dropped WS frames under send failure, duplicated/stale tool cards from deferred-arg regressions, stuck Codex turns after timeout/exit, mid-turn model corruption, silent Windows DPAPI setup desync, duplicate assistant output from abort/rate-limit mis-gating, and signed-runtime trust caching that must not accept incomplete/foreign attestations.

Deliberately avoids duplicating open #298/#299/#300 seams (integrity mid-verify/mid-hash, MCP collisions, idle eviction, canvas route conflict, Responses history, deferred defaults, etc.).

Validation

• Targeted coverage suites + platform-boundary ratchet green
• Touched files Biome-clean (biome check)
• bun run lint, bun run typecheck, bun run docs:check clean
• Full bun run test: all 672 files passed (Bun 1.4.0-canary)