feat: add structured recall evaluation and execution traces

[FloretKu]

Summary

• add versioned, goal-aware JSON recall output with score explanations and correct pre-limit type filtering
• add deterministic offline recall evaluation with Recall@1/3/5, MRR, nDCG@5, leakage/no-result rates, latency, run manifests, snapshots, and run comparison/Goal lift
• add append-only execution traces, judge/human feedback, blockers, validated result states, and review-only Wiki/Skill proposals
• replace the stale frontmatter contract, add formal dataset/trace/manifest schemas, remove model credentials from the core default config, and strengthen health checks

Safety boundaries

• recall evaluation is read-only and fails if persistent KB state changes
• trace payloads reject credential-like fields
• proposals only write to drafts/proposals/; they never modify formal Wiki/Skill content
• the example dataset is format-only and is not presented as human-labeled evidence

Verification

• 43 passed
• Python compileall passed
• git diff --check passed
• wheel build passed; wheel contents include evaluation/trace/proposal modules and all three new schemas

Still required for experiment claims

• 30-50 human-reviewed recall cases
• external Judge/model decision
• fixed GitHub task set, sandbox repositories, execution model, and token budget for the four-way ablation