feat: add comprehensive eval suite for Lore's five key dimensions

[BYK]

Summary

End-to-end eval framework measuring Lore's five key dimensions: context management, multi-session recall, user preference recall, cross-project learning, and cost. Includes 16 scenarios with 130 questions, 6 baselines, and full CI integration.

Infrastructure (packages/core/eval/)

• Two execution modes: fixture (deterministic, free, for CI regression) and live (real LLM calls, for quality benchmarks)
• LLM backend abstraction (llm-backend.ts): Anthropic (preferred), GitHub Models API (free fallback), OpenAI — auto-detected from environment with proper retry/backoff
• Multi-metric rubric judge (judge.ts): 15 scoring criteria, 9 pre-built rubrics, 1-5 scale per criterion with weighted composites
• 6 baselines (baselines.ts): tail-window, compaction+tail, raw, lore context-only ablation, lore memory-only ablation, auto-memory (external Python sidecar)
• Gateway integration: live mode auto-starts an isolated Lore gateway, replays session transcripts to build Lore state (distillation, LTM, recall), then asks questions through the gateway
• Independent cost verification (cost-verifier.ts)
• CLI entry point (run.ts): dimension/baseline/model selection, JSONL output, summary reports

Scenarios (16 total, 130 questions)

Dimension	Scenarios	Questions	Key Tests
Context Management	CM-1/2/3	35	Early detail retention, tool output dedup, layer escalation
Multi-Session Recall	MSR-1/2/3	33	Sequential dev, deep history, cross-model recall
Preference Recall	PR-1/2/3	22	Explicit prefs, implicit patterns, preference evolution
Cross-Project	CP-1/2/3	14	Gotcha transfer, architecture patterns, pref consistency
Cost	COST-1-5	26	Tracking accuracy, short/long/multi-session cost

CI (.github/workflows/eval.yml)

• Fixture mode on PRs touching core/gateway/eval code (deterministic, free)
• Live mode via workflow_dispatch or weekly schedule (requires ANTHROPIC_API_KEY secret)
• auto-memory baseline job (optional, continue-on-error)
• Results uploaded as artifacts

First Live Results (Preference Recall, Sonnet 4.6)

Scenario	Lore	Tail-Window	Notes
PR-1 (explicit prefs)	3.6	4.9	Captures ~60% of explicit preferences
PR-2 (implicit prefs)	4.9	4.8	Lore beats tail-window — LTM adds value
PR-3 (pref evolution)	1.8	5.0	Recalls old pref, not the update

Key findings:

• Lore's LTM/recall adds real value for implicit behavioral patterns (PR-2: 4.9 vs 4.8)
• Curator captures ~60% of explicit preferences from short sessions — room for improvement
• Preference evolution is the biggest gap: curator doesn't supersede old entries when user changes preference
• Cost per preferences-only run: ~$0.42; full suite estimate: ~$5