fix: restore call graph relationships and Steps effects

[colbymchenry]

This restores relationships lost between npm 1.6.0 and main, fixes the remaining Steps/callers assertions, and avoids repeatedly scanning each source file while looking for bound store actions. Excalidraw's mutation-to-render flow connects again, and Steps retains database effects and selected Zustand actions.

• Restore typed JS/TS/TSX field calls, destructured and typed store actions, direct React Native bridge calls, and WASM Dart extension-type getters.
• Retain qualified external call sites in both extractors without guessing internal targets. Resolve proven Zustand selector bindings, including renamed selections and closure captures, while rejecting shadowed bindings and unrelated factories.
• Correct the callers-truncation fixture to two real definitions and assert both markers. Preserve exact precision and native/WASM parity assertions.
• Canonicalize installer test roots for macOS path aliases and isolate the PPID watchdog test's project directory from an editor's writer lock.
• Remove the name-matcher/import-resolver cycle through the resolver context; cover store actions through re-exports and renamed imports.
• Cache per-file .getState eligibility in a bounded resolver-context memo, invalidated with source caches during sync. Positive matching still uses every existing scope, shadow, import and store guard. No helper bypass or reference removal.

Validation on Node 24.16.0/Linux:

• Earlier follow-up passed 233 native resolver/regression cases, 42 affected WASM cases, 245 installer cases (three existing platform skips), and 15 watchdog cases. The final cache follow-up adds three real-SQLite native/WASM edit-and-sync tests covering both eligibility transitions, independent projects and retained reference coordinates/multiplicity. Deliberately removing invalidation makes the negative-to-positive test fail.
• Latest native focused run: 234/236 passed at standard limits; two Objective-C cases exceeded five seconds, including an isolated retry. Baseline Objective-C passed 4/4; optimized passed all four assertions with a diagnostic 15-second allowance during a run with 75 seconds of test collection. This is not a clean default-limit native-suite verdict. No committed timeout or assertion was weakened.
• Latest affected WASM checks verified 45 cases: 32 initially, plus all 13 Steps cases passing in isolation after a combined-run setup timeout.
• TypeScript and asset copying passed. The first build reached its 240-second limit during viewer bundling; the remaining viewer build and all 29 grammar checks passed separately.
• Full Excalidraw before/after indexes have identical node, edge and retained-reference rows/multiplicities, with clean integrity/FK/orphan/error checks: 693 files, 12,779 nodes, 54,045 edges and 38,044 references. Earlier Express/Gin/Django and render-flow evidence is preserved.

Four fresh full VS Code platform indexes and all integrity/FK/orphan checks now complete with no wall timeout. Every node, edge and retained-reference row/multiplicity is identical across both pairs (75,767 nodes, 256,523 edges, 166,516 retained references; zero pending).

CPU configuration	Index CPU before → cached	Index elapsed before → cached
One core, forced sequential	61.33 → 57.58s	144.19 → 133.86s
Two cores, automatic workers	62.73 → 58.21s	126.31 → 128.05s

The cache saves 6.1–7.2% of total indexing CPU in these pairs; resolution CPU falls 8.4–10.8%. Elapsed improvement remains variable. Diagnostics show substantial disk/page waits with only 92–158MiB available at sampled minima, zero visible CPU quota throttling, and 0.21–0.30% CPU steal. Matching CPU does not rise with progress. The old 180s stops came from the audit harness; all incomplete evidence is preserved, and its precise wait sites cannot be reconstructed.

The portable benchmark now records phase/batch progress, CPU/RSS and event-loop state, identifies database maintenance, saves completed-index measurements before verification, and rejects existing fixture indexes. The Linux observer records process/thread and host resource counters without a wall cutoff. Actual native/WASM harness checks preserve a real cross-file call and a qualified external reference; existing-index refusal leaves SQLite bytes unchanged, and a wrong backend produces a nonzero observer exit. No additional resolver behavior changed.

Mac execution remains unavailable here; earlier extreme parser and monolithic-worker limitations are documented. No npm release is included.

Audit and cache follow-up report · Portable indexing harness

Remote head 4e6dcf9bd2b09e38aa0da7b0358ff5fdb084897d has the exact tested local tree 191a128a5c41db20cfef226a1a0b764666f962e4. Original evidence and the unmodified pre-optimization build are preserved locally.