AI-assisted; verified with the tests and tools listed above.

Manual real-app comparison, using the e2e engine only as a scripted browser driver.
No assertion result is used as visual proof.

Base commit: 42185a6206245c1cde5368e8ce368b384e7a400a
Head commit: 534cc7212595229466b0c310028284d0f86c264a
Browser: Chromium 156, Linux, 1600x900, light theme, reduced motion.
Both run fresh servers with T3CODE_DEV_SEED=0; no live database is used.

To reproduce with these files downloaded to one directory:
1. Create disposable checkouts of both commits and run vp i in both.
2. Run python3 prepare-evidence.py /path/to/head-checkout.
3. In head-checkout/scripts, run:
   bash /path/to/run-evidence.sh /path/to/base-checkout base
   bash /path/to/run-evidence.sh /path/to/head-checkout head
4. Inspect images and observations under /tmp/factory-pr-23251-evidence-3.
   Recordings are in scripts/.e2e/evidence-{base,head}/results.

The driver seeds a disposable git project and a completed replay turn through
T3's CLI. It waits for Publish repository, collapses workspace details, renames
the project directory, opens a terminal, and captures the actual terminal error.
It restores the directory, creates a fresh terminal, and runs printf 'RECOVERED\n'.
It reopens workspace details and waits for Publish repository before capturing
recovery. Each final image must match the preceding image taken 500ms earlier.
Terminal text is read through the terminal's own Select All / Copy behavior.

The .mp4 files are continuous excerpts of those browser recordings, beginning
before workspace details are collapsed and ending after recovery. Setup and
pairing are excluded. No app content was modified for captures.
The base and head recovered PNG files are byte-for-byte identical.
The PR changes only the error message; it adds no serialization test.
