feat(kotlin): index @Composable as component, record annotations

[aabolfazl]

Annotation-driven Kotlin was unqueryable: codegraph_search kind=component returned zero, and @HiltViewModel / @Dao / @Entity left no trace on the graph. Framework annotations live in libraries outside the index, so the unresolved decorates ref was the only artifact and it never resolved.

What changed

@Composable functions and methods get the component kind, and annotation names persist onto the node and show in codegraph_node:

$ codegraph query "" --kind component
component   NiaApp        app/src/main/.../ui/NiaApp.kt:89
component   NiaButton     core/designsystem/.../component/Button.kt:50
...  151 components indexed          # 0 on main

$ codegraph node NewsResourceDao
**NewsResourceDao** (interface)
**Annotations:** `@Dao`               # nothing on main

Three extraction bugs fall out of the same collector — all silent before:

Shape	main	this PR
@Preview(showBackground = true)	dropped (constructor_invocation never unwrapped)	Preview
@Dao interface / @Serializable enum class	dropped (path collected no decorators)	Dao / Serializable
@[Suppress("x") JvmStatic]	Suppress only	both

Opt-in per language via LanguageExtractor.extendedAnnotations, which only Kotlin sets: each of the 13 kernel walkers owns its decorator logic, so widening the shared engine alone breaks parity elsewhere — and collecting past the first target is Kotlin-specific, since Swift's @Siblings(through: Pivot.self, from: \.$left) would harvest self and $left as names.

Kotlin routes to the kernel, so codegraph-kernel/src/kotlin.rs and the wasm walker both change; annotation collection moved ahead of node creation in both, because a node's kind and decorators are fixed when it is built.

Blast radius. A component node silently drops out of gates keyed on function/method. Widened alongside: flushFnRefCandidates/defined_fn_names (a real dropped-edge bug — register(::Header) lost its function_ref, pinned by a test), captureValueRefScope/value_scopes, enclosingScopeStartLine, matchFuzzy's callable kinds, the function-ref candidate filter, findDeadCode's defaults, and kmpKindsCompatible for an asymmetrically annotated expect/actual pair.

Tested

Kernel↔wasm sweep, fresh --depth 1 clones. Before is a worktree at main with its own dist/ and its own kernel built from that tree's kotlin.rs, so deltas are attributable:

Repo	Files	Diffs	Deferred before → after	Nodes before → after	Components
okio	327	0	24 → 24	6,764 → 6,764	0
okhttp	617	0	51 → 51	16,396 → 16,396	0
kotlinx.coroutines	1,082	0	51 → 51	15,379 → 15,379	0
nowinandroid	350	0	11 → 11	5,603 → 5,603	151
compose-samples	380	0	16 → 16	8,836 → 8,836	545

0 diffs on 2,756 files; deferral and node counts unchanged (the +1/+2 vs the recorded 23/49/51 is repo drift — the before-tree reproduces 24/51/51 on these checkouts). The three non-Compose libraries correctly produce zero components while still gaining annotation coverage (okhttp 0 → 3,895 annotated symbols).

31 tests in __tests__/kotlin-annotations.test.ts, every guarantee asserted on both arms (CODEGRAPH_KERNEL=0 for wasm, tryKernelExtract for the kernel) — a bare extractFromSource only exercises the kernel for Kotlin. torture.kt extended with the new shapes. Full suite: 3,254 passed, 0 failed.

Not run: an agent A/B. An earlier eval found census questions ("how many @Composable…") don't make the agent reach for codegraph — 3/3 runs chose grep. The component kind and the codegraph_node line ride tools the agent already calls; neither has a measured retrieval win yet. Recorded in docs/design/kotlin-annotation-extraction.md.

Node ids hash the kind, so an existing Kotlin index needs a re-index before component nodes appear.

🤖 Generated with Claude Code