PR #24019 live verification, 2026-10-28, Linux Boat VM
baseline github/main: 9bc4138923 (separate /home/user/main worktree)
PR exact head: 9923e5b2e6fea248828bc08554acb0628eb01b50

Setup:
  git worktree add ../main github/main
  (cd ../main && timeout 600 vp i)
  T3CODE_DEV_SEED=0 setsid nohup vp run dev > /tmp/main-dev.log 2>&1 &
  main actual ports: server 16733, web 8693, state /home/user/main/.t3
  stopped main by its captured process group before PR startup.
  git fetch github refs/pull/24019/head
  git switch -c maria/takeover-pr-24019-verify FETCH_HEAD
  timeout 600 vp i
  T3CODE_DEV_SEED=0 setsid nohup vp run dev --home-dir /home/user/t3code/.t3 > /tmp/dev.log 2>&1 &
  PR actual ports: server 13773, web 5733, state /home/user/t3code/.t3

Client/provider path on both revisions:
  repo dependency Playwright 1.60.0 + /usr/bin/google-chrome (1440x1000).
  paired the web client, used Add project > Local folder > /home/user/terminal-proof.
  this is a real git repo, initialized with an empty initial commit.
  selected Codex gpt-6.1-sol in the composer and submitted:
    run pwd and git status --short in this repo, then reply terminal verification ready. do not change files.
  both real provider turns completed with: terminal verification ready.
  opened the terminal drawer and typed the following workload through its input:
    node -e 'process.stdout.write("a".repeat(8*1024*1024));let n=0;const t=setInterval(()=>{process.stdout.write("b".repeat(1024));if(++n===16){clearInterval(t);process.stdout.write("\nFLOOD_COMPLETE\n")}},600)'

Actual terminal-log syscalls:
  sudo strace -f -yy -s 0 -e trace=openat,write,writev,pwrite64 -p <actual server PID>
  measured only writes to .t3/userdata/logs/terminals/*.log.
  main final 16 flushes: 16 x 8,388,608 = 134,217,728 bytes, O_TRUNC.
  PR final 16 flushes: 15 x 1,024 + 1,093 = 16,453 bytes, O_APPEND.
  final PR flush includes FLOOD_COMPLETE plus the shell prompt.
  PR active file after workload: 8,405,378 bytes; replay snapshot: 8,388,608 bytes.
  this measures terminal-log write amplification only, not CPU, latency or whole-app throughput.

Compaction through the same live terminal:
  node -e 'let n=0;const t=setInterval(()=>{process.stdout.write("c".repeat(512*1024));if(++n===32){clearInterval(t);process.stdout.write("\nCOMPACTION_COMPLETE\n")}},400)'
  syscall trace showed snapshot rewrites after file sizes 16,270,004 and 16,252,928 bytes.
  largest completed file size in trace: 16,270,004 (< 16,777,216).
  after final compaction: 8,388,608 bytes, ending COMPACTION_COMPLETE + shell prompt.

Cold recovery through real server/client reconnect:
  generated another 8 MiB + 16 small updates through the terminal input.
  retained file before abrupt server stop: 8,405,330 bytes.
  killed only the captured dev-server process group, restarted with the same explicit home.
  browser reconnected and terminal.attach returned a new PTY PID.
  restored snapshot: 524,278 bytes, including the existing terminal input-reset sequence.
  the 524,160 bytes before that reset exactly matched the suffix of the pre-stop file.
  FLOOD_COMPLETE was preserved. The existing cold replay window is <= 512 KiB.

Clear and subsequent output:
  authenticated WebSocket RPC terminal.clear(threadId, terminalId) returned Success.
  the persisted terminal log was exactly 0 bytes after the response.
  typed printf '\nAFTER_CLEAR_名_🚀\n' in the same terminal, then reloaded the browser.
  terminal.attach returned 119 bytes including exact UTF-8 AFTER_CLEAR_名_🚀.
  neither FLOOD_COMPLETE nor the old repeated output returned.
  unicode text persisted correctly; this Linux terminal font displays the rocket as a missing glyph.

Existing focused tests (remote Blacksmith worker):
  vp i && vp test run apps/server/src/terminal/Manager.test.ts
  1 file passed, 46 tests passed.
  includes partial-append failure recovery, clears, split Unicode and cold bounded replay.
  worker stopped after completion.

No source edits, commits or pushes were needed.
Not exercised: native desktop/mobile, macOS/Windows, relay/tunnel, other providers,
real disk-full/permission-error/power-loss conditions, or whole-app performance.