perf(filtering): split METADATA into thin + fat tables (v2 schema)

[vlasky]

Summary

Splits the single METADATA table into a thin table (~50 bytes/row) and a fat METADATA_CONTENT table (large TEXT columns), linked by a stable monotonic _content_id_. Re-sequencing after deletes now touches only the thin table, making performance position-independent.

Benchmarks (20K rows, 27KB each, delete near position 0):

• v0 (INTEGER PRIMARY KEY): 940ms
• v1 (perf(filtering): fat-row-independent metadata delete re-sequencing (#136) #141, demoted PK): 211ms
• v2 (this PR, thin table): 8ms - 26x faster than v1, 117x faster than v0

Design

• METADATA retains _subset_, _content_id_ FK, and small filterable columns (file, name, line, unit_type, language, complexity, booleans)
• METADATA_CONTENT holds large TEXT columns (code, signature, docstring, parameters, calls, etc.) keyed by _content_id_ INTEGER PRIMARY KEY
• _content_id_ is stable and monotonic - never re-sequenced on delete
• PRAGMA user_version = 2 marks the new layout
• Orphaned content rows are cleaned up on delete

Backward compatibility

• v0/v1 read and write paths preserved unchanged - existing indices work with no migration
• INDEX_FORMAT_VERSION bumped to 2 in colgrep, triggering a full rebuild on upgrade (no in-place migration needed)
• Removed the v0->v1 in-place migration path; format mismatches now trigger full rebuild

Changes

• next-plaid/src/filtering.rs: v2 create, update_v2, delete_v2, get_v2, update_where_v2, JOIN-based where_condition/where_condition_regexp for fat column queries
• next-plaid/src/text_search.rs: FTS rebuild()/update_rows() adapted to JOIN for v2
• colgrep/src/index/state.rs: INDEX_FORMAT_VERSION: 1 -> 2
• colgrep/src/index/mod.rs: simplified format mismatch handling to full rebuild

Related

• Closes perf(filtering): split METADATA into thin index + fat content table #136 (thin/fat split proposal)
• Builds on perf(filtering): fat-row-independent metadata delete re-sequencing (#136) #141 (demoted PK - the v1 approach this supersedes for new indices)

Test plan

• 130 unit tests in next-plaid pass (including 6 new v2-specific tests)
• 619 tests in colgrep pass
• 54 tests in next-plaid-api pass
• Integration tested on 3,722-file real-world index (search, incremental update, format-version-triggered rebuild)
• Stress tested: delete at all positions, concurrent readers+writer, adversarial inputs (1MB blobs, Unicode, NUL bytes, SQL injection attempts, out-of-range IDs)