Port upstream search, thumbnail, geometry, and cache improvements

[ewlarson]

This selectively ports shared backend improvements reviewed through geobtaa/api 0.9.1. Encoded facet values now survive parsing, date-range-only records gain index years, CONTENTdm/OSU manifests resolve through the correct image services, and static maps/viewers handle projected geometry and points correctly.

The PR also adds opt-in AND faceting across search, facet values, and H3 maps, with separate cache keys and preserved facet links. Existing OR semantics remain the default. Optional Redis response compression uses an OGM version marker and leaves HTTP bodies, ETags, TTLs, and durable records unchanged; compressed writes are disabled by default.

Scope and compatibility

• Adapts upstream eb89c10, 131fa98, f6e6573, 4389011, 4ca8e5f, 9ee88fe, 80dcd59, and 5b35609, with regression tests.
• Preserves OGM Level 0 IIIF handling, endpoint-owned thumbnail queueing, durable coverage state, harvesting, identity, and basemap provider.
• Retains existing app.identity version reporting. Dependency updates, Boolean grouping, visibility policy, institutional fields, and production configuration remain separate work.
• Records the complete review and implementation provenance in docs/upstream_reconciliation.md, docs/upstream_review_2026-09-12.md, and docs/upstream_ports_0.9.1.md; the full-import baseline is unchanged.

Validation

• Full backend suite in disposable Python 3.11 containers with ParadeDB 0.18.11, Elasticsearch 9.0.0, Redis 7.4.6, and OGM's locked FastAPI/Starlette versions: 1,919 passed, 82 skipped, 1 non-strict xpass.
• Application coverage: 62.87%, above the 50% gate.
• make lint-check, make identity-check, and git diff --check pass.

Rollout notes

Reindexing repairs existing year-facet documents; persisted metadata requires re-ingestion or a separate backfill. Static-map cache variants advance automatically, while affected resolved/failed thumbnails need bounded regeneration. Enable compression only after all readers support the codec, and account for existing compressed entries before rolling back to older readers. No production reindexing, regeneration, or deployment is performed by this PR.