feat(eval): user-facing `lore eval` command (recall@k/MRR, per-query misses, one-flag A/B, no-spend)

[BYK]

Summary

Add a user-facing lore eval command that wraps our existing eval suite (packages/core/eval/) and emits recall@5 / recall@20 / MRR with per-query diagnostics and a one-flag A/B toggle, as a no-spend pure read.

Motivation

Our eval suite is already deeper than competitors on methodology (400K-token early/mid/late recall via cm1.eval.ts, tail-window/compaction/raw baselines, objective retrieval scoring separated from the LLM judge #1186, degradation signals #1185, offline no-spend A/B #1202). What we lack is tight, user-facing CLI ergonomics.

The kossvat/mnemonic mnemonic eval command demonstrates the packaging we want:

• Prints recall@5/@20/MRR for a query set plus per-query lines (✓ = recall@5 ≥ 0.5, · = miss).
• Misses print the actual top-k the retriever returned — so you see why it missed.
• A single flag (--no-graph-hop) toggles one retrieval knob for a clean A/B.
• touch_access: false pure-read path so eval never mutates production rankings (no last_reinforced_at / access-count side effects).

Proposed scope (v0)

• lore eval [--file queries.jsonl] [--json] [--ab <knob>] reading a query set (ship a small generic public set + support a gitignored local set).
• Output: recall@k + MRR for the set, per-query pass/miss lines, and returned top-k on misses.
• No-spend pure read: route recall through a non-mutating path so re-running eval does not touch knowledge_meta.last_reinforced_at or any injection/access bookkeeping.
• One-flag A/B toggle for a single retrieval parameter (RRF weight / graph-hop-equivalent / vector-vs-FTS fusion), reusing feat(eval): offline no-spend context-rot A/B from recorded results (#961) #1202's offline harness where possible.

Guardrails

• Coordinate with feat(eval): real long-session + multi-session lore A/B harness (#961) #1211 (real long-session A/B harness) and feat(eval): offline no-spend context-rot A/B from recorded results (#961) #1202 (offline no-spend A/B) — wrap the existing suite, do not fork it.
• Keep retrieval-quality metrics separate from any end-task/LLM-judge scoring (feat(eval): objective retrieval scoring, separated from the LLM judge (#961) #1186), consistent with the Publish a coding-agent memory benchmark (no competitor measures this) #961 no-justifier-inflation rule.

Part of #961.

Filed from competitive analysis of kossvat/mnemonic.

[BYK]

Clarifying scope so this isn't closed as "already shipped": PR #1211 added a lore eval command, but that command is the end-task benchmark reproducer (Lore-on vs Lore-off live counterfactual → fact-retention/efficiency/cost scorecard, per #961). It does not emit recall@k/MRR or per-query miss diagnostics.

This issue (#1237) is a distinct retrieval-quality mode — recall@5/@20/MRR over a query set, per-query miss lines showing the returned top-k, a one-flag A/B knob toggle, and a no-spend pure-read path (no last_reinforced_at/access side effects). Same command name, different measurement. Still open; not delivered by #1211. Tracked under #961.