Fix screenshot eval assertion to match the tool's JPEG output

[lezama]

Related issues

None.

How AI was used in this PR

The code change and this PR description were drafted with Claude Code; the diagnosis was confirmed by running the eval suite locally.

Proposed Changes

The screenshot-all-timing agent eval has been failing because its assertion still expected PNG screenshots, while take_screenshot now produces JPEG (screenshot-desktop.jpg / screenshot-mobile.jpg, image/jpeg). The agent behavior was correct all along — one fast viewport: "all" capture of desktop and mobile — but the check rejected the payload on the file name/format and parsed dimensions as PNG.

This updates the assertion to validate the real invariant rather than the encoding:

• Keys the expected width by viewport (desktop → 1040, mobile → 390) instead of the file extension, so it no longer breaks if the format changes again.
• Reads pixel dimensions from either a PNG or a JPEG buffer.

The eval suite is green again, so the daily eval reports an accurate signal instead of a standing false failure.

Testing Instructions

1. From eval/, run the screenshot case:
   npx promptfoo@0.121.4 eval -c promptfoo.config.yaml --no-cache --filter-pattern "screenshot verification"
2. Confirm it passes (it captures one viewport: "all" screenshot and validates the desktop/mobile JPEG dimensions).

Pre-merge Checklist

• Have you checked for similar PRs that may already address this?
• Verified locally (the case passes).
• Reviewed by a teammate.

[wpmobilebot]

📊 Performance Test Results

Comparing a907ccf vs trunk

app-size

Metric	trunk	a907ccf	Diff	Change
App Size (Mac)	1334.62 MB	1330.69 MB	3.93 MB	🟢 -0.3%

site-editor

Metric	trunk	a907ccf	Diff	Change
load	1633 ms	1634 ms	+1 ms	⚪ 0.0%

site-startup

Metric	trunk	a907ccf	Diff	Change
siteCreation	9053 ms	9032 ms	21 ms	⚪ 0.0%
siteStartup	4427 ms	4420 ms	7 ms	⚪ 0.0%

---

Results are median values from multiple test runs.

Legend: 🟢 Improvement (faster) | 🔴 Regression (slower) | ⚪ No change (<50ms diff)

[sejas]

Thanks for updating the eval 🙌