[L6] Inline preview encoding (Pillow downscale → WebP)

[Seretos]

#ai-generated

Summary

Produce a small, context-budget-friendly preview of an image asset for inline display in an MCP tool result: downscale with Pillow and encode to WebP under a size cap, with a graceful "too large → metadata only" fallback.

Footprint (new/changed files)

• src/lib_python_comfy/preview.py:
  • encode_preview(image_bytes, *, max_dim=512, max_b64_chars=100_000, quality=70) -> PreviewResult (raw WebP bytes + base64 + whether it fit the budget)
• tests/test_preview.py — uses a generated in-memory test image (no network).

Acceptance criteria

• Output respects max_dim and stays under max_b64_chars; quality steps down if needed.
• Non-image / undecodable input fails cleanly with a clear error (caller can fall back to metadata).
• Pure function over bytes — no client/network dependency.

Dependencies

• Must come after L1 (Pillow dep) and L5 (operates on asset bytes / Asset). blocked_by set for both.

Notes

Mirrors encode_preview_for_mcp / view_image in the reference. The MCP-side view_image tool (wrapper ticket A4) calls this.

[Seretos]

#ai-generated

Bundled in feat/m1-finalize-bundle (with #11, #12, #13, #14, #15) — deliberately processed together as one PR since the tickets are too small to slice individually.

This ticket's slice — inline preview encoding:

• New src/lib_python_comfy/preview.py: encode_preview(image_bytes, *, max_dim=512, max_b64_chars=100_000, quality=70) -> PreviewResult.
• PreviewResult (frozen dataclass): webp_bytes, b64, fit. Pillow downscale to max_dim, iterative quality step-down to meet the base64 budget; over budget → fit=False (no exception). Undecodable input → ValueError. Pure, no network.
• Exported from __init__.py __all__ (finalized in [L11] Finalize public API surface + README #11).
• New tests/test_preview.py with in-memory generated images.

Implemented first in the bundle (foundation for #15's WebP optimize). Full plan drives one combined implementation + review + draft PR.