-- 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;