perf(tools): cache GraphStore connection per db_path

[gzenz]

Summary

Cache GraphStore connections per db_path so MCP tool calls don't re-open SQLite on every invocation. Measurable latency win for sequential MCP workflows.

Stacked PR approach

This is PR 1/16 in a stack that ports closed #158 into small, reviewable slices based on your feedback. Each branch contains this commit plus every earlier branch in the stack, so it applies cleanly on its own. The intended merge order is bottom-to-top.

Stack (merge order):

1. feat/store-connection-cache-v2
2. feat/hardened-tests-v2
3. feat/parser-python-star-imports-v2
4. feat/parser-resolve-class-method-v2
5. feat/parser-cross-file-class-method-v2
6. feat/parser-cross-file-class-method-fuzzy-v2
7. feat/parser-node-decorators-v2
8. feat/parser-import-map-bare-call-resolver-v2
9. feat/query-dedupe-and-transitive-tests-v2
10. feat/incremental-ignore-build-artifacts-v2
11. perf/build-timing-telemetry-v2
12. feat/parser-call-noise-filter-v2
13. feat/parser-typed-var-enrichment-v2
14. feat/parser-func-ref-enrichment-v2
15. feat/cli-quiet-json-and-enrich-v2
16. feat/cli-refactor-dead-code-v2

After you merge PR N, I rebase PRs N+1..16 on current main and force-push. Each PR ends up as a single commit on top of the previous one. Review only the newest commit in this PR; the rest will already be on main (or in earlier PRs in the stack).

All 16 branches are rebased on 072ab80. None re-attempts the closed lang-handler refactor.

Roadmap: restoring v7 loop-test parity

Measured impact on loop-test fixtures (Gadgetbridge Kotlin/Java, HealthAgent Python/TS):

• HealthAgent CALLS resolution: 15.2 percent to 36.1 percent (v7 baseline 39.3 percent)
• HealthAgent dead-code: 275 to 122 (backend/api 106 to 0, backend/cli 57 to 2)
• Gadgetbridge CALLS resolution: 33.4 percent, Q2/Q3/Q5 scorecard queries now return qualified targets
• Gadgetbridge loop-test scorecard: 8 PASS / 2 PARTIAL

Still deferred after this stack

• Lang handler refactor (closed feat: lang handler refactor + parser improvements #246): will redo per-language, one PR per handler, each with tests, once v7 parity holds. Addresses the no-handler-tests concern directly. Svelte, TESTED_BY, Zig, PowerShell, and Julia support will be preserved.
• test_pain_points.py fixtures: half are handler-dependent, so the full suite waits for the per-handler PRs.

[tirth8205]

Closing this as incompatible with current main. At head cbe6d7b, the proposed global GraphStore cache conflicts with the current per-tool ownership contract: tools close their stores, so the cache can retain closed connections, and concurrent requests can share a connection that another request closes. The branch is conflicting and its Python 3.11 job failed. A future optimization needs a concurrency-safe lifecycle design plus a benchmark; this branch should not be rebased or merged.