PR: https://git.cafe/pingdotgg/t3code/pulls/29105
Head: 705fe548de4de3b5b6ede8a967c0c4d23f285597
Main: 60aec28a9c
Browser: repo Playwright 1.60 + /usr/bin/google-chrome, real web client
Provider: Codex gpt-6.1-sol, low reasoning, local proxy
Main real app: created project /home/user/t3code via Add project > Local folder.
Sent 400-word snapshot explanation; received STREAM_COMPLETE.
Enabled Labs > Plan mode (legacy); used Shift+Tab.
Asked provider to call request_user_input with Alpha/Beta.
Selected Alpha and Submit. Provider acknowledged Alpha and emitted QUESTION_COMPLETE.
Browser-only response instrumentation counted full projector passes and row reads without changing reducer/provider/output.
Main cumulative counters: calls=48; rowReads=207; requestReads=17.
Main also displayed an existing 'Runtime request ... is resolved' error after answer submission, despite completing the turn. Investigate parity on PR.

PR isolated home: /home/user/t3code/.t3; web 5733; server 13773.
Exact source head checked out; vp i completed on main and PR.
All scenario data created through Add project, composer prompts, and provider question submission. No replay provider, fixtures, mocks, or synthetic rows.
PR: same prompts and same real Codex gpt-6.1-sol path completed.
Pending Alpha/Beta question displayed; selected Alpha, clicked Submit; provider acknowledged Alpha and emitted QUESTION_COMPLETE.
Compared a later identical 350-word follow-up prompt with one saved request on both revisions.
Main sampled projector passes: 21; full scans 21; total 208 row reads and 21 request-index reads.
PR sampled projector passes: 16; includes thread switching and snapshots; 6 zero-full-scan passes. Totals are not a latency benchmark and not directly comparable update counts.
Comparable 10-row replacement: main scanned=10, indexed=1; PR scanned=0, indexed=0. The PR still projects the replaced row. Counters only count full-loop iterations and request map entries.
PR question creation and resolution each triggered a full fallback; stream replacements afterward bypassed full scans.
Neighbor checks: switched away and back; reloaded stock client after removing browser-only instrumentation; Alpha, QUESTION_COMPLETE and INCREMENTAL_COMPLETE all remained present. No page errors on PR.
Main showed a resolved-request error after submitting an answer, then completed. This was not observed on PR. It is not attributed to this projector change.
Verification setup correction: first PR startup used root checkout's default ~/.t3; stopped immediately after detecting it, no test requests sent, restarted with explicit isolated home. Subsequent PR verification used isolated home only.
Blacksmith: vp test run packages/client-runtime/src/state/threadRequests.test.ts packages/client-runtime/src/state/threadDetail.test.ts: 32 passed.
Blacksmith: vp run check:changed c65f8d552baca1b51e15b5fb4737f07e18cce177: all steps passed. Worker needed depth-2 PR history to compare with its base.
Not exercised: native mobile, Electron shell, non-Codex providers, remote/relay/tunnel, 1000-row scale or user-visible latency improvement.
No source changes, commits or pushes.