feat: enhance Ekko skill matching and learning with unified JEV settings

[EKKOLearnAI]

Summary

• Add optional JEV semantic skill matching to Ekko: retain exact name/keyword matches, evaluate bounded metadata candidates, and load up to three confident additions through the existing skill tools.
• Add a JEV preflight to background skill learning. Only a confident negative skips the full reviewer; uncertain results, provider failures and timeouts preserve the existing flow. Queued reviews retain their originating run configuration and cancellation signal.
• Expose both stages as one default-off, Profile-scoped Use JEV for Ekko skills switch in Models → JEV, with shared candidate limit, confidence and timeout controls. Saved settings take effect on the next run. Register the integration, add all locale strings, and migrate standalone configuration to schema 14 without enabling it automatically.

Validation

• npm run test:coverage -- tests/ekko-agent tests/server/jev.test.ts tests/server/jev-harness.test.ts tests/server/ekko-agent-manager.test.ts — 599 tests passed.
• npm run test:e2e -- tests/e2e/jev-settings.spec.ts --workers=2 — 4 tests passed, including desktop/mobile settings persistence and Profile switching; inspected the mobile screenshot.
• npm run harness:check
• npm --prefix packages/ekko-agent run check
• npm run build
• git diff --check

Limits

Semantic routing evaluates a bounded candidate window, while the full skill-name catalog remains available to the main model. Context estimates remain local. Learning preflight falls back for oversized evidence rather than judging a truncated transcript. Provider accuracy, latency and net cost have not been benchmarked on real tasks.