PR 16317 real-environment verification, 2026-11-01
main: 7195b69954 (github/main, separate /home/user/main worktree)
PR: 61b561a16ed09b6f68e3d1ff23217a9bc603341c (exact contributor head)

Both: vp i; background vp run dev; one dev server at a time.
main: web 8693 / server 16733, isolated .t3.
PR: web 5733 / server 13773, explicit --home-dir /home/user/t3code/.t3.
Client: repository Playwright 1.60.0 with /usr/bin/google-chrome.
Created stream-verification project through Add project, pointing at a real git repository.
Selected Codex gpt-6.1-sol in the composer; ran real turns through the local proxy.

Provider command on both heads:
node -e 'const chunk=Buffer.alloc(64*1024,120);for(let i=0;i<256;i++)process.stdout.write(chunk);process.stdout.write("\nSTREAM_BURST_END\n")'
Observed in expanded tool output: Earlier output not shown; STREAM_BURST_END.
Next command: printf 'NEIGHBOR_COMMAND_OK\n'; git status --short
Observed: NEIGHBOR_COMMAND_OK; no changed files.
Assistant response: all 120 requested numbered lines and STREAM_VERIFICATION_COMPLETE.
Reload: completed response retained on both heads.
No live-buffer overflow warning on either head.

PR neighboring check:
Started real Codex turn with a 90-second heartbeat command.
Clicked Stop generation while Running node was visible.
Observed: Run interrupted / Stopped by user; composer returned.
Submitted follow-up: printf AFTER_STOP_OK; reply AFTER_STOP_COMPLETE.
Observed: AFTER_STOP_COMPLETE; retained after reload.

Blacksmith checks:
PR: vp test run ThreadLiveEventCoalescer.test.ts LiveStreamBudget.test.ts: 25 passed.
PR: vp run check:changed c2dc2eaaf5e66cf9ebebb49af7ca8fa7bd41597e:
format, lint, knip, graphql, affected typechecks and 26 tests passed.
Main's coalescer is identical at c2dc2ea and github/main 7195b69954.
Temporarily substituted that implementation in the remote worker, keeping PR's existing regression:
handles a tool output burst larger than the default 8 MiB buffer: FAILED
LiveStreamBufferError: The live event buffer is full. Resume from the last received sequence.
Restored PR implementation afterward. No local repository source changes.

LIMITATION: original overflow was not reproduced through the real app.
Codex command output bypasses this coalescer via CommandOutputHub and is stripped from thread wire events.
The observed PR websocket stream had 8 turn-item.updated events, including 2 running command items,
but no running assistant-message snapshots. No repeated superseded snapshot burst was observed.
Synthetic regression failure/pass is separate from runtime smoke evidence.
Native desktop/mobile, other providers, and relay/tunnel connections were not exercised.