docs: correct OpenAI vision request count

[promiseeuler]

Summary

While reviewing the current OpenAI vision how-to against its runnable example, I found two documentation mismatches. This change:

• adds the example’s explicit version="responses-contact-sheet-v1" to the guide’s decorator;
• clarifies that the shipped example makes one model request per check, or two per episode, while retaining the documented 0.5 FPS sampling and 12-tile cap.

Why

The runnable example has two checks and each independently calls client.responses.create. Describing that as one request per episode understates API usage, while omitting version= conflicts with the guide’s advice to version external model configuration.

Fixes #73.

Validation

• git diff --check — passed
• Repository lychee --no-progress --include-fragments ... command — 260 total links, 257 OK, 3 expected exclusions, 0 errors

Checklist

• No business logic changed, so an outcome-focused test is not applicable.
• I updated the affected documentation.
• Python formatting, typing, and pytest gates are not applicable to this Markdown-only change.
• I ran the repository’s prescribed Markdown link check.
• I did not add recordings, generated media, credentials, private URLs, or runtime artifacts.
• Stored-data compatibility is unchanged.

[kstonekuan]

LGTM. Merging.

Verified against the example: two checks, each with its own client.responses.create, fps=0.5 and max_tiles=12 unchanged, and OPENAI_ENDPOINT_ALIAS = "vision" so the snippet's uses="vision" stays truthful. Copying describe_activity's version string rather than hand_visibility's is the right one for that snippet. Lychee here matches your numbers: 260 links, 0 errors.

"caps each request at 12 tiles" instead of "the request" was not in the issue and is the more correct wording once there are two.

#107 is still sitting closed from the PR cap and can be reopened whenever.