PR 35026 real-environment verification, 2026-11-15 Base: github/main 843893e58c Head: cbf57430576ca96f2826e22933c389cb8c152961 No source changes, commits, or pushes. Commands and runtime path $ git worktree add ../main github/main $ cd ../main && timeout 600 vp i $ setsid nohup timeout 7000 vp run dev > /tmp/main-dev.log 2>&1 & [dev-runner] source=worktree serverPort=16733 webPort=8693 baseDir=/home/user/main/.t3 $ git switch -c fix/projection-unchanged-index-writes cbf57430576ca96f2826e22933c389cb8c152961 $ timeout 600 vp i $ setsid nohup timeout 6000 vp run dev --home-dir /home/user/t3code/.t3 --no-seed > /tmp/dev.log 2>&1 & [dev-runner] serverPort=13773 webPort=5733 baseDir=/home/user/t3code/.t3 Main stopped before PR testing. An initial PR startup selecting the shared home was stopped before sending turns; verification used the explicit isolated home above. Used repo Playwright with /usr/bin/google-chrome, actual web pairing, Add project -> New project (GitHub publishing unchecked), actual Codex turns. No demo data, mock provider, recorded events, or projection writes by the verifier. Main gpt-6.1-sol: requested 24 labeled output lines, proof.txt creation and readback. Observed final output projection-ok, completed run, completed command items, idle provider thread, completed provider turn. PR gpt-6.1-sol: same prompt completed, 24 lines persisted, projection-ok read back. An earlier PR turn used gpt-6-astra because a fresh recording context reset the model; it was repeated with gpt-6.1-sol and the persisted modelSelection was checked. Live database proof Read-only snapshots of real server SQLite WALs, 4096-byte pages. Mapped index roots using sqlite_schema. Parsed complete frames in the current WAL generation, rejecting frames with obsolete salts. Compared successive page bytes, rather than timing a synthetic upsert or replaying events. Index: orchestration_v2_projection_provider_turns_thread_idx Main captured turn: 5 frames, 4 byte-identical rewrites. PR captured gpt-6.1-sol turn: 1 frame (insertion), 0 byte-identical rewrites. Index: orchestration_v2_projection_provider_turns_thread_ordinal_idx Main: 5 frames, 4 byte-identical rewrites. PR: 1 frame, 0 byte-identical rewrites. This proves removal of redundant index writes on a real user turn. It is not an end-to-end latency or generalized throughput claim. Full-upsert status changes can still rewrite other unchanged indexes by design. Neighboring behavior on PR - Follow-up in the same thread completed using gpt-6.1-sol. - One real provider-native gpt-6.1-sol child read proof.txt, completed and returned projection-ok with trailing newline; subagent projection and UI result persisted. - A real timeout 40 shell command printed streamed-neighbor-1 through streamed-neighbor-20, one per second, and exited successfully. All 20 output lines persisted in the completed turn item. - Browser reload retained messages, tool results, child result and completed status. - PRAGMA integrity_check returned ok. Checks $ blacksmith-check run --id [task-owned worker] 'npm exec --yes --package=node@26.13.1 --call "node --version && vp test run apps/server/src/orchestration-v2/ProjectionStore.test.ts"' v26.13.1 Test Files 1 passed (1) Tests 59 passed (59) Initial Node 24 run: 58 passed, 1 failed because StatementSync.stat was unavailable; rerun on required Node 26 passed all 59. Existing forge checks observed: 45 successful, 1 failing (Agent context budget). No code/check fixes were needed for the behavior under verification. Not verified Native desktop Electron shell, iOS/Android clients, macOS/Windows, non-Codex providers, relay/tunnel modes, large-database throughput, and app-managed completionDelivery state transitions. Provider-native child completion was exercised; specialized completionDelivery merge behavior was covered by the existing tests rather than forced through the UI.