Issue 37935 — real server / CLI timeout proof

Base: 2b2d65d131332d9fb15f5390d2d79ab6e21e3825
Head: 78192459742f24428a2995f75f04773fd3851a12
Environment: Linux, Node, isolated fresh T3 database and synthetic git project per side.
No tests or mocked T3 service methods are used. T3 starts the actual Claude SDK subprocess boundary.

Reproduction:
1. Install each checkout with vp i.
2. Run reproduce.py base or reproduce.py head with PROOF_CHECKOUT pointing to that checkout.
3. Driver puts a synthetic claude executable first on PATH and launches the real server.
4. Driver uses project add and thread start --provider claudeAgent --model claude-haiku-4-5.
5. The stand-in answers initialize, deliberately withholds the first mcp_set_servers reply,
   and answers registration and the prompt on any subsequent process attempt.
6. Capture thread show at 5, 95, 110 and 245 seconds. The real MCP timeout is 90 seconds.

The separate startup capability probe (initialize/get_usage/get_settings) is not a turn attempt.
The first mcp_set_servers request identifies the thread's first attempt.
No Claude OAuth account is required. The synthetic response verifies delivery after recovery,
not real model quality or live Claude service behavior.

The MP4 files are timed playbacks rendered from captured protocol messages and sampled real CLI
JSON, at 10x elapsed time. They are not browser or terminal-screen video captures. Status is
explicitly labeled with its sampling time. The PNG is the final frame of the same playback.
The transcripts retain the actual timestamps and CLI state. Rendering does not determine verdicts.

Result: base fails at the 90-second timeout and never retries through 245 seconds; head retries
at 90 seconds, delivers the prompt to attempt 2 and completes the original run.
