# PR 23548 evidence-only, round 2

Head: 67d2406b222abda6222b97560ee878008c1518f0
Independently derived merge-base: 43879d970204eb28092d098deb4c842a22a9ab75
No implementation changes, commits, pushes, merges, comments, assignments, or PR watching.

| Row | Result | Evidence / limitation |
| --- | --- | --- |
| Real web runtime | Partial | Head fixture rendered simultaneously: row 27/30, panel 3 of 30 running; these agree. No progression captured. |
| Before/after progression | Unverified | Before source was never substituted; no divergence/settled comparison. |
| Recordings and measured narration | Blocked | Recording transfer returned a path under prohibited ~/.t3/userdata/attachments. Artifact not opened/copied; deleted by exact path. |
| 1440 light / 390 dark still pairs, two diffs each | Unverified | Only initial head pending 1440x1000 light screenshot available from inline image bytes. It contains a provider-update toast. No pair or diffs. |
| Server cache timestamp hit/refresh | Test evidence only | Existing focused PullRequestService tests passed; no runtime RPC timestamp trace. |
| Duplicate reference RPC count | Test evidence only | Existing shared query tests passed; no actual WS request count captured. |
| Old server / new client | Unverified | Reviewed fallback and existing resolver tests; no integrated mixed-version capture. |
| Merged PR + open checks | Test evidence only | Existing resolver tests passed; no integrated rendered case. |
| Cross-server clock skew | Unverified runtime; review finding | Reviewer A FIX_FIRST; reviewer B MERGE, same issue rated low. No two-origin offset capture. |
| Mobile | Source review only | Shared detail query normalization applies. Native views do not use the web checks polling/resolver. No native parity claim. |
| Desktop | Source review only | Wraps web; no Electron-specific pass. |

209 existing focused tests passed (3 files), 34.89 seconds. No new tests.
CI read via cafe: 32 successful, 0 pending, 0 failing; conflict readiness pending native check in stored status.
What else changed visually: no comparison available; initial screenshot includes synthetic title/author/body, extra default server project, elapsed thread labels, and provider-update toast.

Blocker: app-owned screenshot persistence and recording transfer write into ~/.t3/userdata. Screenshot bytes were separately obtained inline and saved under /tmp without reading the forbidden file. Recording exposes only the forbidden local path; no alternative browser permitted by test-t3-app. The recording prerequisite cannot be completed under the supplied path constraint. No real T3 database was read, copied, or opened. Only fresh synthetic /tmp state was accessed.

Driver SHA256:
- provider-driver.mjs: d9369f0cc3552a6ade7fbc92fc3d2c3107a453a2fa3da577e9a65c4b44259154
- seed.py: a54098be374fede0434663ffc8de104b4d72075089d6e4acd5c1a976c2510d81

FRICTION:
- fixture | Initial driver used passing instead of contract success; corrected before partial capture | Validate fixture against contract first.
- dev watcher | Vite test config creation restarted server; replacement owner recorded | Run focused tests before starting dev.
- browser evidence | Artifact paths are inside prohibited T3 home; no recording bytes exposed | Configurable evidence output directory outside live T3 home.
- reviewer | Claude ran a redundant detached checkout despite read-only brief; no HEAD or file change | Avoid checkout in shared evidence worktrees.

Model/harness: primary gpt-6.1-sol / Codex in T3 Code; reviewers gpt-6.1-sol / Codex and claude-opus-5-5 / Claude Agent.