fix(capture): skip oversized SQLite rows instead of aborting source

[robbe1912]

Summary

Provider SQLite databases can contain data values larger than ctx's 16 MiB per-value import cap. Before this PR, materializing one oversized value could abort the entire source import with SQLITE_TOOBIG, leaving otherwise readable rows unindexed.

This keeps the cap and makes the OpenCode-family SQLite adapter omit oversized values at the SQL layer instead of aborting the source:

• pre-scan relevant tables with a read-only, query-only connection that returns row ids only;
• exclude oversized ids from the bounded import query before SQLite materializes data;
• keep a per-row SQLITE_TOOBIG safety net for rows that slip past the prescan;
• count omitted rows as skipped / skipped_events, not failures;
• continue importing every other readable row.

No synthetic text, fake previews, truncated diffs, or summary events are generated. Oversized payload content is omitted.

Affected scope: OpenCode-family SQLite rows currently read by normalize_opencode_sqlite, including OpenCode/Kilo dialects. JSONL adapters are unchanged.

Validation

• cargo fmt --check
• cargo test -p ctx-history-capture --lib tests::native_sqlite
• cargo test -p ctx-history-capture

Regression coverage includes mixed readable + oversized rows, all-oversized rows, and legacy message table rows.