clip-search: migration + client + probe binary

Probe-phase scaffolding for CLIP semantic search. Adds the column
that will hold per-photo embeddings, the HTTP client to Apollo's
inference service, and a throwaway probe binary so we can eyeball
search-result quality on the live library before building the
persistence layer (backlog drain, /photos/search endpoint, UI).

- migrations/2026-05-14-000000_add_clip_embedding/ — adds
  image_exif.clip_embedding (BLOB) and clip_model_version (TEXT),
  plus a partial index on (clip_embedding IS NULL AND content_hash
  IS NOT NULL) for the future backfill drain.
- src/database/models.rs — extends ImageExif struct to match.
- src/ai/clip_client.rs — encode_image / encode_text / health,
  same Permanent/Transient/Disabled taxonomy as face_client.
- src/bin/probe_clip_search.rs — --query <q> --library N --limit M
  --top K. Encodes a sample and prints top-K cosine similarities.
  No DB writes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

migrations/2026-05-14-000000_add_clip_embedding/down.sql

@@ -0,0 +1,3 @@
+DROP INDEX IF EXISTS idx_image_exif_clip_backfill;
+ALTER TABLE image_exif DROP COLUMN clip_model_version;
+ALTER TABLE image_exif DROP COLUMN clip_embedding;

migrations/2026-05-14-000000_add_clip_embedding/up.sql

@@ -0,0 +1,27 @@
+-- CLIP semantic photo search: store a per-photo image embedding so
+-- text queries can rerank against the live library via cosine
+-- similarity. Apollo encodes the bytes via its CLIP service; ImageApi
+-- writes the resulting blob here.
+--
+-- `clip_embedding` is the raw little-endian float32 buffer of an
+-- L2-normalized vector (dim depends on the model — 768 bytes×4 for
+-- ViT-L/14, 512 bytes×4 for ViT-B/32). Apollo always returns the
+-- normalized form so the search-time dot product reduces to a plain
+-- cosine similarity.
+--
+-- `clip_model_version` echoes the upstream `APOLLO_CLIP_MODEL` (e.g.
+-- "ViT-L/14"). A model swap shouldn't silently mix geometries — the
+-- backfill drain will re-eligibilize rows whose stored model_version
+-- differs from the live engine's, and the search route refuses to
+-- mix rows from two model_versions in the same response.
+ALTER TABLE image_exif ADD COLUMN clip_embedding BLOB;
+ALTER TABLE image_exif ADD COLUMN clip_model_version TEXT;
+
+-- Partial index for the backfill drain. Mirrors the shape of
+-- `idx_image_exif_date_backfill`: candidate rows are those with a
+-- known content_hash (so we don't race the unhashed backlog) but no
+-- embedding yet. SELECT cost stays O(missing rows) instead of full
+-- table scan once the column is mostly populated.
+CREATE INDEX IF NOT EXISTS idx_image_exif_clip_backfill
+    ON image_exif (id)
+    WHERE clip_embedding IS NULL AND content_hash IS NOT NULL;