[spelech]

Fixes #77.

Changes

• Replace both SQLite-only PDF extraction-cache writes with one shared INSERT ... ON CONFLICT ... DO UPDATE statement supported by SQLite and PostgreSQL.
• Preserve extraction fields on cache refresh without aborting PostgreSQL metadata transactions.
• Add actual HTTP/file/database regressions for text and rasterized PDFs, upload/replacement, persisted cache reads when OCR is unavailable, stale-cache refresh and deletion.

Verification

• Before the fix: 4 native PostgreSQL cases failed while 4 SQLite controls passed.
• PDF extraction and cache cases: 21 passed, 1 skipped.
• Generated requirement inventory: 2 passed; release/link-integrity gate passed.
• Combined backend regression run: 70 passed, 1 skipped.
• With the companion backend fixes: 2 full PostgreSQL/pgvector/API-key-auth profile checks passed, including injected vector-write failure and successful retry.

The isolated cache regression uses real PostgreSQL metadata and embedded Qdrant with an external synthetic embedding/OCR HTTP endpoint. The combined profile additionally verifies real pgvector persistence/search. Full-suite and fresh-source-build acceptance are not claimed.