feat: wire adaptive analysis runtime into cache and review flow

[wimi321]

Runtime Integration

• Adds analysis/runtimeIntegration.ts and routes scheduled KataGo IPC through runtime wrappers.
• Keeps runScheduledAnalysis as the outer scheduling and priority layer.
• Runtime integration now owns adaptive profile selection, cache lookup/write, evidence bundle attachment, and teaching readiness evidence.
• This branch includes V1 and V2 foundations because current main does not yet contain them.

Performance: adaptive profile + cache

• Adds main-side analysis cache keys using game/model/config/tier/visits/move/schema version fingerprints.
• Supports sweep, refine, teaching, and oracle tiers.
• Adds cache statuses: hit, miss, stale, lower-quality, schema-mismatch, corrupt, written.
• Adds adaptive analysis profiles based on speed mode, benchmark visits/sec, and task intent.

Accuracy: classification / PV confidence / readiness

• Adds moveClassification and pvConfidence to KataGoMoveAnalysis without removing legacy fields.
• Adds teaching evidence bundle and readiness gate.
• Adds optional runtimeEvidence with cache status, adaptive profile, readiness level, safe wording, warnings, blocking issues, and evidence bundle version.
• Teacher prompt now explicitly respects runtimeEvidence.teachingReadiness and cache status.

UX: review runtime model

• Adds timeline review item model, review session model, PV preview/playback helpers, teacher evidence panel model, and combined runtime review model.
• These are pure helper/model layers for later App/board/timeline visual integration.

Verification

• pnpm install --frozen-lockfile
• pnpm test
• pnpm typecheck
• pnpm build
• pnpm check:teacher-quality