pr 23445 verification, linux boat vm, 2026-11-01

main: 5732bbee397f0a2b0e61d0c2e432f01dca6df3dc
pr:   3fa7145f7c9234f331349d75b18aca099711cfb6

scope: test-only replay temporary-directory cleanup. normal app turns do not
import these testkits. replay lifecycle evidence and live-app smoke evidence
are separate checks. no code changes or new tests were made.

replay checks ran through blacksmith-check after vp i inside each worker.
each run used its own initially empty TMPDIR, counted after the process exited.

main adapter command:
vp test run apps/server/src/orchestration-v2/Adapters/{Codex,Claude,AcpRegistry,Grok,Fx}AdapterV2.test.ts
result: 5 files, 447 tests passed; 117 shared replay directories remained.

main integration command:
vp test run apps/server/src/orchestration-v2/testkit/{AcpRegistry,Grok,Fx}ReplayFixtures.integration.test.ts
result: 3 files, 40 tests passed; 80 directories remained:
40 shared, 15 acp registry, 16 grok, 9 fx.

pr command: both commands above plus
apps/server/src/orchestration-v2/testkit/ProviderReplayHarness.test.ts
result: 9 files, 483 tests passed; LEFTOVER_COUNT=0.
the lifecycle test confirms stateDir exists while scoped and baseDir disappears
after scope closes. integration scenarios complete with their replay status
files alive until teardown, covering success, interruption, and error paths.

live app, one dev server at a time, fresh isolated data, no fixture/demo data:
main: server 16733, web 8693, /home/user/main/.t3
pr: server 13773, web 5733, /home/user/t3code/.t3

created a real project through the running server:
node apps/server/src/bin.ts project add --base-dir <isolated-home> --title 'pr 23445 verification' /home/user/t3code

started a real provider turn through the running server on both revisions:
node apps/server/src/bin.ts thread start --base-dir <isolated-home> --provider codex --model gpt-6.1-sol --title 'real codex verification' --wait --json /home/user/t3code 'Run pwd and git rev-parse --is-inside-work-tree. Do not modify any file. Reply with the two results.'
both: status completed, error null
reply: pwd: /home/user/t3code
       git rev-parse --is-inside-work-tree: true

repo playwright-core with /usr/bin/google-chrome drove the actual web app.
pr composer follow-up: Without changing files, run git rev-parse --short HEAD and reply with the result.
reply: 3fa7145f7c
reloaded the page and verified both provider replies remained visible.
recording shows submit, real provider processing, tool result and final reply.

not exercised: native desktop/mobile clients, macos/windows filesystem cleanup,
real claude/grok/fx/acp agents, full repository suite. these test helpers have
no production user-facing route; the live codex check is neighboring behavior.

vp run check:changed through blacksmith-check: passed format, lint, knip,
graphql, typecheck (3 packages), tests (2 files). worker needed depth-3 PR
history and the PR merge-base ref to compare the five changed files.