PR #27405 real-environment verification, Linux Boat VM, 2026-11-01
main: github/main 5732bbee397f0a2b0e61d0c2e432f01dca6df3dc
PR: 264e9f9bb57ba100ce4999fb6f11538fd76f9e12 (exact contributor head; no edits)

Setup:
- git worktree add ../main github/main; vp i
- Main: background vp run dev --no-seed, web 8693 / server 16733
- Stop captured main process group before starting PR
- PR: vp i; background vp run dev --no-seed --home-dir /home/user/pr27405-state
- PR web 5733 / server 13773
- Playwright from repo dependencies, /usr/bin/google-chrome
- Projects created through the web client against real git directories
- Actual Codex gpt-6.1-sol turns, no mocks or seeded database

Original unnecessary serialization reproduced on main:
- Complete Codex turn: assistant says "replay verification ready."
- Disconnect web UI while creating backlog through real orchestration.dispatchCommand RPC
- 128 thread.metadata.update commands: 127 distinct long titles, final short title
- Stored raw payload: 1,043,887 bytes, below 1,048,576-byte raw guard
- Full encoded replay: 1,087,791 bytes
- Node inspector breakpoint on existing JSON.stringify(item) line: 128 hits
- Resume from pre-backlog sequence 177: snapshot at 305, then synchronized
- Snapshot preserves provider answer

Exact PR head:
- Same user flow and 128 rename commands, title size calibrated to raw cap
- Stored raw payload: 1,043,933 bytes, below raw guard
- Full encoded replay: 1,101,860 bytes (older head has longer thread identifiers)
- Node inspector breakpoint on existing JSON.stringify(item) line: 122 hits
- Resume from pre-backlog sequence 42: snapshot at 170, then synchronized
- Six trailing items skipped; snapshot preserves title and provider answer
- No breakpoints during latency sampling; inspector detached before UI recording

Neighboring behavior on both heads:
- Resume from last sequence minus one: one thread.metadata-updated event, then synchronized
- Resume from a future cursor: snapshot, then synchronized

Real web reconnect and continued provider use on PR:
- Temporarily offline, close existing app WebSocket, restore connection
- Client issues orchestration.subscribeThread afterSequence=171 again
- Server emits synchronized and then streams the subsequent turn's events
- Send follow-up through normal web composer using Codex gpt-6.1-sol
- Assistant says "follow-up after reconnect works."
- Original answer, restored title and follow-up remain visible

Existing tests through blacksmith-check:
vp test run apps/server/src/orchestration-v2/ThreadStream.test.ts
7 passed
vp test run apps/server/src/ws.test.ts -t "replay|resume"
4 passed, 5 excluded by name filter
Existing PR CI: 30 successful checks, 0 pending, 0 failing

Not verified:
- Native desktop, iOS/Android, other providers, relay/tunnel and multi-device paths
- Reliable wall-clock speedup: real WebSocket request medians were 22.0ms main
  and 22.6ms PR (20 warm samples each); different head schemas and development
  overhead prevent attributing a timing improvement to the small early exit
- No source changes, no fix commits or pushes, no new test files