fix(extract): dry-run reports net-new rows, not raw candidates (closes #397)

[vinsew]

Summary

gbrain extract links|timeline --source db --dry-run reports the candidate count, not the would-be net-new row count. On the second run after a real extract has populated links / timeline_entries, every candidate already exists in the DB, but dry-run still reports them all as "would create N." ON CONFLICT DO NOTHING in addLinksBatch / addTimelineEntriesBatch would silently no-op the lot on a real run.

This is the problem #397 set out to fix. That PR landed before v0.32.8's multi-source threading (#860) reshaped both extractors, so the old patch can't be cherry-picked onto master:

• Candidate dedup keys now carry source ids (6 segments: from_source_id::from_slug::to_source_id::to_slug::link_type::link_source).
• engine.getLinks() does not return f.source_id / t.source_id, and Link has no source-id fields. Comparing a 6-segment candidate key against a 5-segment DB row key would false-positive cross-source rows.

This PR closes #397.

Approach

Per-extractor inline SQL that returns source ids alongside the link / timeline shape, plus a per-(from-page, source) cache:

async function existingLinkKeysForFrom(fromSlug, fromSourceId) {
  const rows = await engine.executeRaw(\`
    SELECT f.slug AS from_slug, t.slug AS to_slug,
           l.link_type, l.link_source,
           f.source_id AS from_source_id, t.source_id AS to_source_id
    FROM links l
    JOIN pages f ON f.id = l.from_page_id
    JOIN pages t ON t.id = l.to_page_id
    WHERE f.slug = \$1 AND f.source_id = \$2
  \`, [fromSlug, fromSourceId]);
  return new Set(rows.map(r =>
    \`\${r.from_source_id}::\${r.from_slug}::\${r.to_source_id}::\${r.to_slug}::\${r.link_type}::\${r.link_source ?? 'markdown'}\`
  ));
}

The key shape is byte-for-byte identical to the candidate dedup key the extractor already builds (extract.ts line ~872), so the existing/seen sets compose cleanly.

Why inline SQL instead of extending engine.getLinks()

The cleaner-looking alternative is to add from_source_id / to_source_id to Link and to getLinks' SELECT. That ripples to 6 callers outside extract.ts:

• src/core/operations.ts:734 (auto-link reconciliation)
• src/core/operations.ts:1418 (per-page link API)
• src/core/output/validators/back-link.ts:27 (back-link validator)
• src/core/output/validators/back-link.ts:37 (back-link validator)
• src/commands/migrate-engine.ts:234 (engine migration)

None of those need the source ids. Widening the interface across all five callers for a feature only extract --dry-run consumes is the wrong shape. Inline SQL keeps the blast radius at the two extractor functions.

Test plan

4 new cases in test/extract-db.test.ts:

• links: dry-run after a real run reports zero new add_link rows
• links: dry-run reports a newly-added candidate after the page body changes
• timeline: dry-run after a real run reports zero new add_timeline rows
• timeline: dry-run reports a newly-added entry after timeline section changes

Each pair pins both halves of the contract — the dedup must remove "already in DB" candidates AND must not eat genuinely new candidates.

Run results:

$ bun run typecheck
$ tsc --noEmit
(clean)

$ bun test test/extract-db.test.ts
 16 pass
 0 fail
 36 expect() calls
Ran 16 tests across 1 file. [1.79s]

Scope

Unchanged: real-run output, JSON shape, command-line surface, engine APIs.

Only changes:

• src/commands/extract.ts — adds 2 cached helpers (existingLinkKeysForFrom, existingTimelineKeysForSlug) and a 4-line continue guard inside each extractor's dry-run branch.
• test/extract-db.test.ts — 4 new tests.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith with what you need.}

• Let Codesmith autofix CI failures and bot reviews