feat(provenance): persist chunk↔block mapping and page range

[mrsibe]

What does this PR do?

Persists the chunk → block mapping and the denormalized page range, so a retrieval result can be resolved back to its ordered blocks and its page without re-parsing the document.

This is the #68 half of #67, now rebased directly onto main after #106 (the chunking half) merged. It supersedes #107, which GitHub closed when its stacked base branch was deleted.

Why?

A retrieval result only carries a chunk id. Even with correct canonical offsets, provenance cannot survive in the database without the mapping — "chunking without a persisted block mapping is not a partial feature, it is an unverifiable one" (#68, epic #82).

Related issue

Fixes #68. Completes the persistence half of #67.

What changed?

• Schema + additive migration 0015_colossal_solo.sql
  • chunk_blocks(chunk_id, block_id, start_in_block, end_in_block) with a composite primary key and an index on block_id.
  • Denormalized chunks.page_start / chunks.page_end (both NULL for non-paged sources), so a citation does not need a MIN/MAX join on the hot path.
  • Existing documents are untouched and keep working at document-level citations.
• src/main/services/chunkProvenance.ts
  • insertChunkBlocks() writes the mapping.
  • resolveChunkProvenance() resolves a chunk id with a single join (chunks → chunk_blocks → document_blocks) and returns the ordered block spans plus the page range.
  • projectChunkProvenance() is the pure ordering/projection step and is unit-tested.
• KnowledgeService
  • saveChunks() writes each chunk and its mapping in one transaction, so a partially indexed document never has a chunk without its mapping, and stores page_start/page_end.
  • getChunkProvenance() exposes the resolver.
  • deleteDocument() and reindexDocument() delete mappings explicitly, because the connection does not enable foreign-key cascades.
• Packaged smoke test now round-trips the real database: insert a document + its blocks + a chunk + the mapping, then resolve the chunk and check the block spans and page range.

How was this tested?

• npm test — 110 tests pass. test/chunkBlocks.test.ts gains the ordered-span projection and the page-range invariant (page_start/page_end equal the min/max page of the mapped blocks).
• The packaged smoke test covers the database round trip on all three CI platforms, which is the only environment with a working better-sqlite3 (the Node unit-test runner has an Electron-ABI build).
• npm run typecheck — clean. npm run build — clean. npx eslint — clean.

Screenshots / recordings

Not applicable — no UI change.

Checklist

• I have reviewed my own changes.
• npm run typecheck passes.
• npm run build passes.
• I have tested the affected user workflow. (DB path covered by the packaged smoke test in CI)
• I have not included unrelated changes.
• I have updated documentation when necessary.

Desktop / build changes

• Not applicable
• npm run build:unpack passes. (delegated to CI)
• npm run smoke:packaged passes. (delegated to CI; includes the provenance round trip)