fix: code-source first-sync (#744) + class-method code-def

[lanceretter]

Summary

Two atomic commits that together make gbrain sync --strategy code actually work end-to-end on first sync, and gbrain code-def resolve class methods that the existing tree-sitter chunker already extracts.

Discovered while upgrading a real PlanetScale Postgres brain from v0.26.6 → v0.30.0 (companion to #740, #741, #744).

1. commands/sync.ts + commands/import.ts — strategy threading (closes #744)

Reproduces issue #744 directly. performFullSync called runImport(repoPath) without the strategy flag; runImport then walked markdown only via collectMarkdownFiles. First-sync of a fresh gstack-code-* source reported "X pages imported" but persisted zero code files. Three different repos on the same brain showed identical broken state.

Three small changes, all additive:

• Extract a shared collectFiles(dir, strategy) walker in commands/import.ts that branches on strategy. Add collectCodeFiles and collectFilesAuto wrappers. collectMarkdownFiles keeps its current behavior + signature for backward compat (no caller changes needed).
• runImport accepts strategy in opts and picks the right walker. Default 'markdown' preserves pre-strategy callers.
• performFullSync threads opts.strategy through to runImport (and into the dry-run filter at line 861, which dropped strategy too).

Verified end-to-end across three repos:

Repo	Code files imported	Status
conquest-lpr (a 5-app monorepo)	858	✅
rv-helper (Expo + Supabase)	360	✅
trashtastic-helix	partial — 2 legitimate failures (5.6MB minified dist/ JS; Supabase types.ts with NUL bytes — both correctly refused, not bugs)	✅

2. commands/code-def.ts — extend DEF_TYPES for class members

gbrain code-def <symbol> returned 0 results for any class method even when chunks existed with the correct symbol_name, because the WHERE filter only allowed:

['function', 'class', 'interface', 'type', 'enum', 'struct', 'trait', 'module', 'contract', 'export statement']

The tree-sitter chunker emits more granular symbol_type values than that. Real distribution on my brain after a full sync:

function                  | 10897
declaration               |  5926
method definition         |  3201   ← was filtered out
export statement          |   818
import                    |   375
class                     |   343
interface                 |   215
field definition          |   124   ← was filtered out
variable assignment       |   108
struct specifier          |    60
public field definition   |    57   ← was filtered out
type                      |    40
method signature          |    14   ← was filtered out

Concretely: Database.calculateDeviceStatuses (a TypeScript class method) returned {count: 0} from code-def even though the chunk existed at apps/worker/src/db.ts with the right symbol_name and symbol_type='method definition'.

Fix: add four multi-word symbol_types representing class members. Additive only:

   const DEF_TYPES = [
     'function', 'class', 'interface', 'type', 'enum', 'struct', 'trait', 'module', 'contract',
+    'method definition', 'method signature', 'field definition', 'public field definition',
   ];

'declaration' is left out (too generic — covers imports, type aliases, plain bindings).

After: gbrain code-def calculateDeviceStatuses returns 3 hits (the real source plus a couple of compiled-JS build artifacts that are themselves a separate walker concern).

Test plan

• Build: bun install && bun run build — passes.
• Typecheck: bun run typecheck — passes.
• Existing tests: bun test test/schema-bootstrap-coverage.test.ts — 5/5 pass (unchanged).
• Smoke test: fresh gstack-code-* source on Postgres → gbrain sync --strategy code --source <id> actually imports code files; gbrain code-def <classMethod> returns hits.
• CI on this PR — open to suggestions if there's a specific test you'd want for either fix.

Notes on follow-up bugs surfaced during this dig

Not in this PR; flagging for future:

• The walker doesn't respect .gitignore — tmp/, dist/, build artifacts get indexed alongside source. Worth a .gitignore-aware skip-list or an explicit --ignore glob.
• --source <id> is recorded on sources.last_commit/last_sync_at but pages still land under source_id='default' rather than the named source. That's why gbrain sources list shows page_count=0 on the new sources even after a successful sync. Reasonable next dig — happy to PR if you want.

🤖 Generated with Claude Code

---

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

• Let Codesmith autofix CI failures and bot reviews