fix(design): bump image-gen timeout to 240s + pin gpt-image-2

[matteo-hertel]

Problem

/design-shotgun (and any flow using the design binary) appears to hang indefinitely and never opens the comparison board.

Root cause: the binary calls POST /v1/responses with model: gpt-4o + the image_generation tool at quality: "high", 1536x1024, but aborts the request after a hardcoded 120 000 ms (design/src/generate.ts, variants.ts, evolve.ts, iterate.ts).

That class of request consistently takes ~140-160 s end-to-end. Every generation aborts before the image is ready. In /design-shotgun, Step 3c launches N parallel agents each calling $D generate; every one aborts at 120 s, retries (another 120 s), all N fail, the board never opens, so the skill looks like an infinite hang.

Evidence

Reproduced the exact API call the binary makes, with a longer budget:

Call	Result
identical request, 300 s budget	HTTP 200, valid 1.7 MB image, 143.5 s
same with model: gpt-image-2	HTTP 200, valid 1.85 MB image, 143.4 s
binary abort deadline	120 s fires every time

Then ran a real /design-shotgun end-to-end with the patched binary, 3 variants generated in parallel:

Variant	Generate time	Quality gate
A	150.0 s	pass
B	161.0 s	pass
C	152.1 s	pass

The 161 s case is the key data point: a naive bump to 150 s would still have failed it. 240 s leaves real margin while still bounding a genuinely stuck request.

Change

• Bump the AbortController timeout 120_000 to 240_000 in generate.ts, variants.ts, evolve.ts, iterate.ts (both call sites in iterate.ts).
• Pin the image_generation tool to model: "gpt-image-2".

Testing

• design/test/variants-retry-after.test.ts (exercises the changed retry/timeout path): 5 pass, 0 fail.
• Real /design-shotgun run: 3/3 variants generated and passed the vision quality gate, comparison board served.
• The design/test/feedback-roundtrip.test.ts failures are pre-existing and unrelated: they fail identically on clean main (session.clearLoadedHtml is not a function, a browse-module breakage in write-commands.ts). Not touched by this PR.

🤖 Generated with Claude Code

[badcom]

There's a related PR open for this same issue: #1528

[matteo-hertel]

There's a related PR open for this same issue: #1528

@badcom the two PRs are fundamentally different. The fact that a PR exists doesn't meant it's the only viable solution.
The one you linked makes the timeout configurable, this one increases the timeout and bumps gpt-image-2.
Ultimately I get little value by having the time configurable, and therefore I made this.
If the other one is merged then feel free to close this, I don't mind at all