perf(ai): cache local model status probes

[JustYannicc]

What changed

• Cache stable Parakeet and Qwen3 local ASR helper status probe results for a short TTL so the AI Settings poll loop does not launch both helpers every second.
• Keep downloading and error behavior compatible, and force-refresh download, warmup, and transcription action paths so explicit flows still read fresh helper state.
• Add focused cache tests with an injected fake slow probe.

Why

• The AI Settings poll loop was launching both local ASR helpers every second even when helper status was stable.
• Audit baseline: 60 synchronous trivial helper launches averaged 1.763 ms each, so the two helper polls cost about 3.526 ms of main-process blocking per second when fast.

Compatibility impact

• downloading and error behavior remains compatible.
• Download, warmup, and transcription action paths force-refresh helper state so explicit flows still read fresh helper state.
• Draft until the integration stack lands. This clean PR branch contains only the status-probe caching fix, and depends on PRs perf(runtime): reduce extension and menubar churn #530-perf(search): reduce indexing and ranking overhead #535 plus codex/perf-consolidated-integration-stack for the broader wave-2 integration context.

How tested

• node scripts/test-local-model-status-probe-cache.mjs
• ./node_modules/.bin/tsc -p tsconfig.main.json
• On the integration worker branch: node scripts/test-ai-model-status-polling.mjs
• On the integration worker branch: pnpm run build:main
• LSP diagnostics for touched TypeScript files passed on the integration worker branch; src/main/local-model-status-probe.ts is clean on this branch. origin/main currently reports unrelated main-process LSP diagnostics for missing/import-mismatched files outside this change.

Performance evidence

• New fake-probe test on the clean PR branch: iterations=60 legacyCalls=60 legacyMs=105.099 cachedCalls=1 cachedMs=1.813 avoidedCalls=59.
• Same test on the integration worker branch before the clean cherry-pick: iterations=60 legacyCalls=60 legacyMs=105.153 cachedCalls=1 cachedMs=1.816 avoidedCalls=59.

Stack validation

• This clean PR branch contains only the status-probe caching fix.
• Depends on PRs perf(runtime): reduce extension and menubar churn #530-perf(search): reduce indexing and ranking overhead #535 plus codex/perf-consolidated-integration-stack for the broader wave-2 integration context.

[JustYannicc]

Superseded by #556, which consolidates this AI/native media performance batch.