PR 34014 runtime verification
Main: 04f6cfca87203192ccf2c36f7ba8c80baf9ba9da
PR: c6903399a8db4ec1d566c266815bf7c1fe7f164c

$ vp run dev (main worktree; T3CODE_DEV_SEED=0)
[dev-runner] serverPort=16733 webPort=8693 baseDir=/home/user/main/.t3
Created project and 13 turns through the web app with Codex gpt-6.1-sol.
One completed turn persisted 160 command_execution items (462024 payload bytes).
Captured before screenshots and recording; stopped captured process group 20878.

$ vp run dev --home-dir /home/user/t3code/.t3 (exact PR head)
[dev-runner] serverPort=13773 webPort=5733 baseDir=/home/user/t3code/.t3
Migration 81_TurnItemInlineWorkOutput applied to copied app-created state.
Read-only SQL comparison against the original LIKE predicate:
rows=218 inline=1 mismatches=0

Real browser navigation on main and PR:
GET /api/orchestration/threads/<thread>/bounded?compactTurnItems=1&dedupeBodies=1&compactIds=1&deferWork=1 -> 200
GET /api/orchestration/threads/<thread>/history?...&deferWork=1&compactIds=1 -> 200
Both older pages returned deferredWork.itemCount=160.
Expanding work requested three history pages with view=work; all returned 200.
Expanded command list rendered in the real UI; inline HTML remained visible.
Full transcript endpoint preserved all 218 items across migration.
Two orphan command statuses were terminalized during normal shutdown.

Fresh PR Codex gpt-6.1-sol turn:
git status --short -> exit code 0
html_render -> completed; inline_work_output=1
Reload -> Post-migration provider turn and POST_MIGRATION_COMPLETE visible.
Read-only SQL: rows=224 inline=2 mismatches=0

$ blacksmith-check run --id <task-owned-box> 'npm exec --yes --package=node@26 -- vp test run apps/server/src/persistence/Migrations/081_TurnItemInlineWorkOutput.test.ts apps/server/src/persistence/reconcileV2PreviewMigration.test.ts apps/server/src/orchestration-v2/ProjectionStore.test.ts'
Test Files 3 passed (3)
Tests 60 passed (60)
Initial Node 24 worker run: 59 passed, 1 required Node 26 SQLite statement.stat.
GitHub checks: passing at the exact PR head; no source edits or pushes.

Limits:
The claimed 2.5x speedup was not established on this small app-created history.
The original large-dynamic-tool-payload slowdown was not reproduced at that scale.
Main's html_preview timed out twice; html_render succeeded on both revisions.
No native desktop/mobile clients, other providers, MCP apps, or delegated-task outputs exercised.
A brief root-checkout startup used the default home; it was stopped and discarded.
All reported PR verification used the explicit isolated .t3 home above.