Stress + agents/vision/cli (terra round, snapshot10g, gpt-6.1-terra/high): 10/13 pass, 3 FAIL (f1, f2, f3) on G-Code defects.
PASS: h1 4 concurrent drivers, 4 valid sessions, no lost write (no same-file race exercised); h2 12/12 in 20 s, no 429/529; h3 8/8 modules all tests pass, no cross-talk; h4 loser refused 'Session … is open in another G-Code process (pid …)', exit 1, no data lost; h5 5 queued turns strictly in order; e1 3 parallel subagents, 21 requests incl. children, $0.33494 exact; e2 22/22 tests, subagent writes in parent outputs; e3 workflow correctly ignored from project settings; i1 fizzbuzz ok, stdout one JSON; i2 background server gone after -p exit.
FAIL f1/f2/f3 [HIGH, G-Code]: src/llm/images.ts:12 visionCapabilities has no gpt-6 entry, so for gpt-6.1-terra context.vision is unset and FileRead refuses PNG/JPEG as 'Binary file detected (.png)'; the Responses input_image path is never reached. Official page: Input modalities text, image. The refusal wording is also misleading for a model without vision.
f3 [MEDIUM, G-Code]: src/browser/observations.ts:63 text observation reads textContent, so <h1>Grown in<br>the mist,</h1> is observed as 'Grown inthe mist,'; the model edited the page to pass (the exact thing shownText was meant to prevent).
f3 not exercised: subfolder entry, bad server.path, 404 URL. Continuations all match; no length stops; no 429.
