PR #26935 real dev verification, 2026-10-30, Linux + Google Chrome

main: 5ef56853ef1854da50bc1909c6c0a32700f2cb13
PR: 1e5d6f49e628a348b8fca69e580b58bcbd9881e7

Setup: separate github/main worktree; vp i on both revisions. One background dev server at a time. Main ports 8693/16733; PR ports 5733/13773. Explicit isolated PR home /home/user/t3code/.t3. Project and thread created through the web app, pointing at the real main git worktree. Codex gpt-6.1-sol through the local provider proxy. No mocks, fixtures, replay provider, source modifications, commits, or pushes.

Real running server, live command event projection:
Temporary RegExp.test counting instrumentation forwarded every call unchanged and recorded stacks from toolOutputIndicatesFailure -> projectTurnItemForWire -> projectDomainEventForWire. Instrumentation removed after each capture.
main: 6 regex calls for one clean 32020-character command result.
PR: 1 regex calls for one clean 32020-character command result.
main regex sources:
file not found|no files found|enoent|no such file|commandnotfoundexception|command not found|is not recognized as the name of a cmdlet|a parameter cannot be found that matches parameter name
cannot find path
is not recognized
<exited with exit code\s+[1-9]\d*\s*>
exit(?:ed)? with exit code\s+[1-9]\d*
exit code\s*[:\s]\s*[1-9]\d*\b
PR regex sources:
enoent|no such|no files found|not found|commandnotfound|not recognized|cannot (?:find path|be found)|exit
This is a scan-count observation through the real provider/server/client path, not an elapsed-time speedup claim.

Main HTTP thread snapshot (200, x-t3-orchestration-protocol: 2):
1. exitCode=0; outputIndicatesFailure=false; /bin/bash -lc "node -e \"console.log('clean-output-26935 ' + 'abcdefghij'.repeat(3200))\""
2. exitCode=0; outputIndicatesFailure=true; /bin/bash -lc "printf 'file not found\\n'"
3. exitCode=0; outputIndicatesFailure=true; /bin/bash -lc "printf 'clean exit first, then ExIt CoDe: 3\\n'"
4. exitCode=0; outputIndicatesFailure=false; /bin/bash -lc "printf 'exit\\nnot found\\nthe path cannot be found\\n'"
5. exitCode=7; outputIndicatesFailure=true; /bin/bash -lc "sh -c 'exit 7'"

PR fresh Codex turn, HTTP thread snapshot (200):
1. exitCode=0; outputIndicatesFailure=false; /bin/bash -lc "node -e \"console.log('clean-output-26935 ' + 'abcdefghij'.repeat(3200))\""
2. exitCode=0; outputIndicatesFailure=true; /bin/bash -lc "printf 'file not found\\n'"
3. exitCode=0; outputIndicatesFailure=true; /bin/bash -lc "printf 'clean exit first, then ExIt CoDe: 3\\n'"
4. exitCode=0; outputIndicatesFailure=false; /bin/bash -lc "printf 'exit\\nnot found\\nthe path cannot be found\\n'"
5. exitCode=7; outputIndicatesFailure=true; /bin/bash -lc "sh -c 'exit 7'"

Expected failure vector for both: false,true,true,false,true. Actual vectors match.
Client command rows also match; failed aria labels only on commands 2,3,5. After page reload and thread reopen the same five labels and flags remain. Expanded actual command output streamed through orchestration.subscribeCommandOutput: file not found\n; clean exit first, then ExIt CoDe: 3\n; exit\nnot found\nthe path cannot be found\n; empty output for exit 7. Large clean output expands and remains clean.

Blacksmith: toolOutput.test.ts 12/12; WireProjection.test.ts 14/14; presentation.test.ts 18/18. vp run check:changed: format, lint, both knip passes, typecheck (31 packages), tests (2 files, 13 tests) passed. Remote shallow history was deepened to include the PR parent so check:changed compared exactly the two changed files.
Existing CI at head: 30 successful, 0 pending, 0 failing.

Not verified: native desktop/iOS/Android; other providers; remote/relay/tunnel transports; Windows PowerShell output through a real Windows provider. No reproduced end-to-end latency or heap ratio; the PR's numeric performance claims remain independently unverified. Existing 32768-character scan limit unchanged and not boundary-tested in this live pass.