feat(celery Wave 6 #37): wire EmbeddingService.embed_image into application cache

[earayu]

Summary

Task #37 (Wave 6, multimodal embedding cache wiring) — ~30 LOC production change wrapping EmbeddingService.embed_image with the canonical aperag.cache.NAMESPACE_EMBEDDING infra (PR #1734). Mirrors the existing text _embed_batch cache pattern.

Wave 5 P2 chunk 4 landed embed_image(image_bytes, alt_text) API; this PR adds the cache layer so identical image embeds short-circuit the LiteLLM provider call.

Cache key shape

Per aperag/cache/README.md no-raw-bytes policy:

{
    "kind": "image",
    "provider": ...,
    "model": ...,
    "api_base": ...,
    "api_key_hash": sha256(api_key),
    "file_hash": sha256(image_bytes),
    "alt_text": ...,
    "multimodal": True,
}

• Image bytes identified by sha256 hex digest → bounded Redis key.
• alt_text part of key (paired text+image inputs return different vectors when hint changes).
• kind="image" discriminates from text embed_query keys in the same embedding namespace.

Tests

New tests/unit_test/llm/test_embed_image_cache.py (7 tests):

• ✅ identical (bytes, alt_text) → second call hits cache (no upstream)
• ✅ same bytes + different alt_text → distinct keys, both compute
• ✅ different bytes + same alt_text → distinct keys, both compute
• ✅ key shape uses sha256 file_hash, raw bytes never appear in key
• ✅ caching=False bypasses cache (always upstream)
• ✅ multimodal=False raises EmbeddingError (defense-in-depth)
• ✅ empty image_bytes raises EmptyTextError

Full unit suite: 1022 passed, 29 skipped, ruff + format clean.

Out of scope

Provider-specific multimodal embedder format variations (Voyage / Jina v3 / OpenAI multimodal SDK input shapes) stay on task #39 per PM dispatch + simple-stable directive.

Test plan

• uv run pytest tests/unit_test/ — 1022 passed
• uvx ruff check + format — clean
• Cache key shape verified via unit test (no raw bytes in key)
• Cache hit/miss + alt_text + bytes-difference behavior pinned