Summary

Adds focused regression coverage for high-risk seams that were still untested on main, while deliberately avoiding duplication of open coverage PRs #298 and #299.

Risky behavior now covered

1. Runtime integrity mid-hash mutation — a tree that mutates after exact hashing but before attestation must fail closed (Runtime changed while it was being verified.) and must not write a trust attestation.
2. Canvas document save conflicts through the workspace JSON-RPC route — conflict results preserve external file content and must not broadcast cowork/workspace/fileChanged.
3. OpenAI Responses history conversion — pending tool calls get synthetic No result provided outputs; cross-model conversion sanitizes/remaps pipe-delimited tool-call IDs, strips thought signatures, and drops sibling-model fc_ item IDs.
4. Windows sandbox setup newest-wins fallback — when marker created_at is missing/invalid, marker mtime decides which DPAPI credential state propagates.

Test files added/updated

• src/coworkRuntime/integrity.ts — tiny test-only afterVerifyExactTree hook (no production behavior change)
• test/coworkRuntime.test.ts
• test/jsonrpc.workspace-route.test.ts
• test/runtime.openai-responses-shared.test.ts (new)
• test/platform/windowsSetupSync.test.ts

Why these tests materially reduce regression risk

• Integrity attestation/caching protects a ~GB signed runtime tree; a miss here can trust a mutated runtime.
• Canvas conflict broadcast suppression prevents clients from treating a failed save as authoritative.
• Responses history normalization is on the critical path for tool-call continuity and API payload validity across model switches.
• Sandbox mtime fallback prevents stale Windows credential state from clobbering newer setup and triggering setup churn.

Validation

• Targeted: 34 pass across touched suites
• Touched files Biome-clean (biome check)
• bun run typecheck / bun run lint / bun run docs:check clean
• Full bun run test: all 673 files passed

Notes

• Open PRs #298/#299 were left alone; this run covers residual gaps outside those diffs.
• No flaky/environment-dependent tests added.