PR 28343 real development verification, 2026-11-01
main: 8b74a69019 (github/main at verification start)
PR: dc48a23a36a89a39d1dae5fc9e86bb204e7f3df4

$ git worktree add ../main github/main
$ (cd ../main && vp i)
$ T3CODE_DEV_SEED=0 setsid nohup vp run dev > /tmp/main-dev.log 2>&1 &
main server 16733; web 8693; task-local home /home/user/main/.t3
Project created with New project in web UI, creating a real local git repo.
Provider selected in composer: Codex gpt-6.1-sol.
First turn: pwd, git status --short, printf TURN_START_MAIN_OK; completed.
Follow-up: 12 separate printf calls; completed with HISTORY_OK.
24 real turn items persisted across the two runs (8 + 16).
The actual turn-start SELECT appears in server.trace.ndjson, successful.
EXPLAIN QUERY PLAN for source SELECT against app-created database:
SEARCH ... USING INDEX orchestration_v2_projection_turn_items_thread_ordinal_idx (thread_id=?)
This reproduces the history-wide access path, not a multi-second delay.
Stopped main's captured process group before starting PR.

$ git switch -c behmot/turn-start-run-item-lookup-20261031 dc48a23a36a89a39d1dae5fc9e86bb204e7f3df4
$ vp i
$ T3CODE_DEV_SEED=0 setsid nohup vp run dev --home-dir /home/user/t3code/.t3 > /tmp/dev.log 2>&1 &
PR server 13773; web 5733; task-local home /home/user/t3code/.t3
Recreated project and same first two prompts through web UI with Codex gpt-6.1-sol.
First turn completed; requested commands ran and returned TURN_START_MAIN_OK (4 persisted command items).
Follow-up completed, 12 command executions, HISTORY_OK.
Reloaded client; messages and command history persisted.
Third turn asked for first turn's marker without tools; returned TURN_START_MAIN_OK.
Created separate thread through New thread; printf and git status completed, SIBLING_THREAD_OK.
All four runs completed, all selected gpt-6.1-sol.
Four successful turn-start query spans captured. PR trace truncates SQL before INDEXED BY.
Therefore EXPLAIN uses the full SELECT extracted from the exact PR source, not the truncated trace.
EXPLAIN QUERY PLAN against app-created PR database:
SEARCH ... USING INDEX orchestration_v2_projection_turn_items_run_ordinal_idx (run_id=?)
USE TEMP B-TREE FOR LAST TERM OF ORDER BY
Read-only comparison of hinted and unhinted query on every real run:
8, 16, 3, 7 items respectively; identical payloads and order.
Missing thread and missing run returned empty results.
No synthetic rows, fixtures, or mocks used for runtime verification.

$ blacksmith-check run --id <task-owned-box> 'vp test run apps/server/src/orchestration-v2/ProjectionStore.test.ts apps/server/src/orchestration-v2/TurnStartReads.test.ts'
47 tests passed, 2 files.
$ blacksmith-check run --id <task-owned-box> '... && vp run check:changed 54c5f25249961a1aa0b61932329092eadbd9d6fc'
Passed format, lint, both knip passes, graphql check, typecheck (3 packages), selected tests.
Explicit merge-base ref used because source snapshot does not contain remote main history.
Dependencies installed on worker with vp i before checks.
GitCafe reports 29 successful checks, 0 pending, 0 failing at tested head.

Limitations: no multi-thousand-item thread or measured speedup, production database,
large tie groups or spill measurements. No native desktop/mobile clients, other providers,
remote/relay/tunnel or multi-device flow. These are not claimed verified.
No code changes, pushes, merges, rebases, force pushes, labels, or other PR mutations.