Persist real face embeddings during actor indexing (unblocks #74)

[ik020]

Prep work for #74 (find people using a reference image).

The bug

Actor indexing detects a face, aligns it, and computes a normalized
SFace encoding per detection — but that encoding was never actually
stored. _actor_records() wrote a placeholder embedding=[0.0] into
every StorageRecord sent to Chroma, so the actor collection's
vector index has never contained real, searchable face embeddings.
Matching during indexing only ever compared against
ActorIndexState.known_encodings, an in-memory list scoped to a
single indexing run, discarded the moment that run finished.

This silently blocks #74: finding a person by reference image
requires comparing a new embedding against every previously detected
face across the whole repository, and there was nothing durable to
compare against.

The fix

• Carry the computed encoding through into each detection dict in
  process_actor_samples.
• _actor_records() now writes that real, normalized encoding as the
  StorageRecord.embedding, instead of [0.0].
• _actor_cluster_records() (the per-cluster summary rollup) is
  intentionally left unchanged — a summary spans multiple detections
  and has no single face image to encode.

Schema version bump — please read

Because this changes what's actually persisted per detection, any
existing generation with the actor modality enabled has
placeholder vectors, not real ones, and is no longer trustworthy for
face search.

This repo already has a mechanism for exactly this situation:
CompletedGenerationManifest.index_schema_version is typed as
Literal[INDEX_SCHEMA_VERSION], and local_snapshots.py's
validate_generation() calls model_validate_json() against the
manifest on disk, explicitly catching ValidationError and
re-raising it as IndexSchemaError. This PR bumps
INDEX_SCHEMA_VERSION from 7 to 8, so old generations now fail
that validation cleanly with IndexSchemaError instead of silently
being treated as complete, searchable indexes with meaningless
vectors.

User-facing impact: any existing generation with actor enabled
will need to be re-indexed after this lands. This is intentional —
the alternative is a search feature (#74) that runs against garbage
data with no indication anything is wrong.

Testing

• New test: test_actor_indexing_persists_real_face_embeddings_not_ placeholders — drives process_actor_samples through mocked
  YuNet/SFace calls end-to-end and asserts the StorageRecord sent
  to storage.upsert carries the real normalized encoding, not the
  placeholder.
• Updated test_actor_records_preserve_stable_detection_metadata,
  which built fake detection dicts without an encoding key; added
  an explicit assertion that the resulting embedding matches the
  real value now that it's used, not [0.0].
• Full suite run clean: 590 passed, 2 skipped, 0 failed.
• Specifically re-ran everything touching INDEX_SCHEMA_VERSION or
  actor indexing (test_generation_manifest.py,
  test_local_snapshots.py, test_indexing.py,
  test_actor_results.py, test_cli.py, test_frontend_app.py) —
  all green, including the existing test that deliberately constructs
  an invalid INDEX_SCHEMA_VERSION + 1 manifest to prove the
  rejection path works.

Not included here

This PR does not add reference-image search itself — that's #74's
actual feature, tracked as a separate follow-up PR once this lands,
since it depends on real embeddings existing at all.

[SaadBazaz]

Hey @ik020,
Thanks for the Pull Request.
As a starter, we require all contributors to "star" and "fork" the repo, in order to determine if there really is a human behind the wheel or an autonomous agent.
Please star the repo for our review to proceed.
Thanks!