PR #23769 real-development verification, 2026-10-28

$ git -C /home/user/main rev-parse HEAD
bd84b1781a (github/main; full SHA available in git)
$ git -C /home/user/t3code rev-parse HEAD
4ee85d950097f9172e5a93f87dad76d83e0d832a

Main was installed with timeout 600 vp i and started in the background.
[dev-runner] mode=dev source=worktree /home/user/main serverPort=16733 webPort=8693 baseDir=/home/user/main/.t3
agent activity publishing standby; waiting for T3 Connect link reconciliation

Playwright 1.60.0 + /usr/bin/google-chrome:
- added /home/user/relay-verify through Add project > Local folder
- selected Codex gpt-6.1-sol and submitted a README tool-read request
- observed: RELAY_VERIFY_OK Relay verification workspace.
- read-only persisted run: model=gpt-6.1-sol, status=completed
Main process group captured at launch: 14276. Stopped before PR startup.

PR was installed with timeout 600 vp i and started in the background:
vp run dev --home-dir /home/user/t3code/.t3 --no-seed
[dev-runner] mode=dev source=default ports serverPort=13773 webPort=5733 baseDir=/home/user/t3code/.t3
agent activity publishing standby; waiting for T3 Connect link reconciliation

Playwright + Chrome, actual project creation and provider turns:
- README tool-read response: RELAY_VERIFY_OK Relay verification workspace.
- reloaded thread and observed persisted response
- follow-up response: RELAY_FOLLOWUP_OK
- reloaded and observed both responses
- read-only persisted runs: ordinals 1 and 2, model=gpt-6.1-sol, status=completed
- a separate initial recording attempt used gpt-6-astra; it is not the shared recording
- shared recording explicitly selects gpt-6.1-sol before submission

Blacksmith remote commands, exact PR head:
$ vp i
exit 0
$ vp test run infra/relay/src/agentActivity/DeliveryAttempts.test.ts infra/relay/src/agentActivity/ApnsDeliveries.test.ts infra/relay/src/agentActivity/MobileRegistrations.test.ts
Test Files 3 passed (3)
Tests 42 passed (42)
$ vp run --filter t3code-relay typecheck
exit 0 (rerun with a 600-second command limit after the first 60-second command limit expired)

GitHub head checks: no failing or pending contexts; several jobs skipped, including relay/web/mobile preview deployments.
GitHub mergeable=CONFLICTING; mergeStateStatus=DIRTY.

Changed-behavior result: UNVERIFIABLE.
No real relay delivery-attempt rows or cron pruning were exercised.
vp run dev starts server/web only. No relay/Clerk/APNs/cloud provisioning credentials
or registered iOS push device were available. Connections has no T3 Connect sign-in.
Ordinary provider turns do not execute the changed Cloudflare cron.
No synthetic rows, mocks, or demo data were used for the running-app checks.
The default main startup copied an existing project with zero threads; verification
used a separately created real git repository and app-created project/turns.

Original bug cannot reproduce on requested current github/main:
commit c3b40b97f7 (#26141) already added retention.
Current main: 7 days, 5000 rows/batch, at most 4 batches/run.
PR head: 14 days, one unbounded delete with created_at index.

Startup correction: the first PR startup briefly used default /home/user/.t3.
It was stopped immediately after reading that log, before scenario data creation.
All PR scenario data was created after restarting with explicit isolated .t3.
No PR source edits, commits, pushes, rebases, merges, label changes, or closures.