PR 15848 maintainer verification — 2026-10-19 UTC
Base: 1451e7cb23034d29774338608368a15d8845df36
Original contributor head: 9a6e730ce7554b9c2cd0f9bee1bc3ca9b21a8fb1

Before/after behavior:
- Main's 283 original store/ChatView logic cases pass. With the minimum new behavior cases, main fails 2 (283 pass); metadata persistence is absent.
- Original PR implementation fails the stale-status/send regression (284 pass, 1 fail).
- Final implementation: 319 unit cases in 5 existing suites and 48 component cases in 3 existing suites pass.
- 8 original added tests in 2 files reduced/rebuilt as 6 added behavior cases in 3 existing files. No new test files.
- First-send lifecycle cases cover dispatch ownership, successful foreground/background helper dispatch, failure/retry recovery, and best-effort link failure.
- Restoring the code reviewed in round 2 makes the closed-details component regression fail: the wrong old PR is linked. Restoring exact fixed bytes passes.
- Web-only typecheck, lint/format on the 16 changed web files, and scoped whitespace checks pass.
- Three complete independent parallel review rounds; latest GPT-6 Sol and Claude Opus 5.5 reviews report no confirmed findings.

Actual app, isolated Linux server and sparse test repository, 1280x800:
1. Chat branch picker -> real GitHub PR lookup -> Local checkout -> receiving draft.
2. PR page Check out -> In this repository -> receiving draft -> ordinary first send accepted -> server projection records:
   branch=fix/link-pr-on-checkout
   PR=15848, repository=pingdotgg/t3code, source=manual
   Full page reload preserves association.
3. Checkout draft with branch confirmed -> close thread details -> terminal switches repository to main -> reload -> ordinary first send accepted -> server projection records:
   branch=main
   pullRequests=[]
4. Earlier unrelated thread remained without a PR link.

Provider is a deterministic local Codex app-server fixture, not a real model invocation. No model calls were billed. The claim is accepted first send and durable link projection, not real agent task completion.

Limits: full checkout/send races, retry and background helper behavior have focused test proof, not separate full-client race recordings. PR handoff task entry point has code/component coverage, not an integrated task run. Remote/relay/tunnel, macOS, Windows, iOS, Android and Electron shell not verified live. Mobile and multi-model fanout are excluded from this PR.