fix(workbench): carry six upstream reliability fixes

[filipgutica]

Summary

Import six merged native T3 reliability fixes into the Workbench fork. These prevent failed SQLite writes, missed long review conversations, disrupted RPC connections, unwanted restart continuations, and staging changes during commit-message generation.

Changes

• #15488: acquire SQLite's write lock before transaction reads and classify native SQLite failures correctly. Read-only connections retain deferred transactions.
• #15427: follow review-reply pages and hold the remark watermark when reads are incomplete.
• #15563: release a full RPC stream buffer when closing the stream so the connection can continue reading replies.
• #15470: recognize relay disconnect messages carrying the network hint while preserving genuine sign-in errors.
• #15604: clean up leftover background work without waking settled native threads after restart.
• #15532: generate commit context using a temporary index; stage the real index only when committing.

Each fix retains its upstream cherry-pick reference. The combined Effect patch retains the fork's existing cache fix; the regenerated lockfile preserves package versions and the existing Alchemy patch. No Workbench-owned implementation, native persistence schema, UI integration, or workflow changes are included.

Verification

• Regression tests reproduced all six defects before the production changes: 19 failing cases. Afterward, all 25 targeted cases pass, including partial staging, split indexes, unborn repositories, and merge metadata.
• The broader focused run passes 740 tests across 14 files, including native restart replay and Workbench store/execution tests. One unchanged baseline failure remains in GitVcsDriverCore.test.ts: preserves renames, unusual paths, modes, and binary statistics.
• Contracts, shared, client-runtime, server, web, and Workbench typechecks pass. The Workbench application-import/dependency-cycle boundary gate passes. Scoped formatting and lint pass with two unchanged lint warnings.
• Against upstream 4ee6bfd50ef4a089440d5c3662db2298da9cc50e, both base 7530667 and head 520f907 have the same existing conflict paths: ChatView.tsx, state/query.ts, and pnpm-lock.yaml. No additional conflict paths or Workbench integrations are introduced.
• To inspect the stream regression, run vp test run packages/client-runtime/src/rpc/session.test.ts -t 'full buffer'. Closing the saturated stream must allow a subsequent server probe to complete.
• Published head 520f9073c9836f53d165ddcbf90028d5fd7511be passes full Workbench CI, quality and desktop smoke, and all three browser-regression shards.
• The private Workbench demo is published for that exact head. Preview verification passes source-container and exported-bundle pairing, HTTP/WebSocket access checks, and crash recovery. The protected launcher and publisher use main 7530667154d5b52adb5c0ac4d9e24f5279659a6f.

Limits

• Writable-connection transactions take the write lock upfront, trading concurrency with other processes for reliable read-then-write behavior.
• The upstream PR-watch fix caches reply tails by comment count. A delete-and-add between sweeps that leaves the count unchanged can still hide the new tail reply until the count changes or the watch restarts. Independent review reproduced this remaining upstream limitation; the previous fork never read these tail pages.
• This cloud workspace's network proxy rejects direct Vercel access with CONNECT 403. Exact-head publication is verified through the successful preview workflow and its revision-pinned PR comment; interactive launcher access from this workspace remains unverified.

Model: GPT-6. Harness: Codex in the saved Linux cloud environment.

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	4.9 KiB	4.9 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.7 KiB	3.7 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.4 KiB	20.4 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	4.9 KiB	4.9 KiB	+41 B (+0.8%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.7 KiB	3.7 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	+41 B (+3.5%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	20.7 KiB	20.8 KiB	+41 B (+0.2%)	29.3 KiB	✅
Claude	Live turn messages	1	2	+1 (+100.0%)	8	✅

Baseline: 7530667 · PR result: 520f907 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[github-actions]

Workbench demo

Open the private 20-minute demo. Sign in to Vercel, then choose Open demo. Reopening this link in the same browser resumes its running demo; Start new demo creates a fresh one. Jira starts disconnected.

Application, runtime startup, fixtures, and native smoke: PR 520f9073c9836f53d165ddcbf90028d5fd7511be. Dockerfile, build configuration, bundle export, publisher, and protected launcher: main 7530667154d5b52adb5c0ac4d9e24f5279659a6f.

Expired? Return to this link and launch again. Use synthetic content with the free OpenCode model.