fix(vision): convert SVG/unsupported image formats to PNG before embedding into history

[JAlmanzarMint]

Summary

vision_analyze could embed an SVG (or BMP/TIFF) image into conversation history with media_type: image/svg+xml. Anthropic's vision API only accepts image/jpeg, image/png, image/gif, or image/webp, so the request fails with a non-retryable 400:

messages.N.content.0.tool_result.content.1.image.source.base64.media_type:
Input should be 'image/jpeg', 'image/png', 'image/gif' or 'image/webp'

Because a vision tool-result is baked into immutable conversation history and re-sent on every subsequent turn, this permanently wedges the session — hermes --resume <id> re-sends the same poison payload and fails identically every time. Retries cannot clear bytes already in the request.

Root cause

_detect_image_mime_type() correctly returns image/svg+xml (and image/bmp) for those file types, but both consumption sites — the native-vision fast path (_native_vision_fast_path) and the auxiliary-API path (vision_analyze_tool) — then base64-embedded the image with that unsupported media_type and shipped it to the provider.

Fix

Add _normalize_to_supported_image(), called before any image is base64-embedded:

• SVG → rasterized to PNG, best-effort via cairosvg → svglib+reportlab → rsvg-convert → inkscape (all soft deps).
• Other non-supported raster formats (BMP, TIFF, …) → re-encoded to PNG via Pillow.
• If conversion is impossible → returns an actionable error (e.g. "install cairosvg or convert to PNG") instead of embedding a session-wedging payload.

Wired into both call sites so the whole bug class is fixed, not just the one site the symptom surfaced at. Temp PNGs are cleaned up on the same path as downloaded images.

Test Plan

• All 99 existing vision tests pass (tests/tools/test_vision_native_fast_path.py, tests/tools/test_vision_tools.py).
• Verified end-to-end on a real SVG: image/svg+xml → normalized to image/png, embeds cleanly.
• Reproduced the original wedge from a real session (media_type 400 on every resume); confirmed the fix prevents an unsupported media_type from ever entering history.

[alt-glitch]

This was generated by AI during triage.

Related (competing approaches, not a duplicate) to the native-image-format cluster: #47299 (Codex data-URL downgrade to text placeholder), #47523 (reject corrupt/unsupported native inputs — same tools/vision_tools.py file but reject vs transcode), and #25936 (gateway-side transcode of uncommon formats). This PR is the transcode-in-vision_tools.py approach wired into both the native fast path and vision_analyze. Flagging the cluster so a maintainer can pick a single mechanism.

[teknium1]

Merged via PR #57890 — your commit was cherry-picked onto current main with your authorship preserved in git history (ac94f2c). Your SVG rasterization ladder (cairosvg → svglib → rsvg-convert → inkscape) and the format-normalization step now run behind the new unified image-source resolver, so SVG/BMP/TIFF sources are auto-converted to PNG before embedding exactly as you designed. Small follow-ups added on top: mkdir for the converted-PNG output dir on fresh installs, stdin=DEVNULL on the rasterizer shell-out (a subprocess guard that landed after your PR was written), and thread-offloading the conversion. Thanks!