feat(tools): build-ahead analyzers for #249 (negatives separation) + #250 (retrieval ablation)

[helebest]

What

Two analysis tools that consume the dikw-core fixes currently in flight, written
against the proposed output so they run the moment #249/#250 land — no code change
needed when they do.

• tools/negatives_separation.py — pos-vs-neg top-1 relevance-score
  separation for an eval NDJSON. Probes the candidate score-field names #249 may
  add (scores / ranked_scores / top1_score / …) so it is name-agnostic;
  degrades to a rank-only observation with scores_available=false (never a bogus
  zero) until #249 ships. Computes the separation margin + expect_none
  satisfaction at a cutoff.
• scripts/ablate_retrieval.py — retrieval-config ablation harness
  (rrf_k / weights / fusion / rerank). Merges each variant's overrides into a
  temp base's retrieval: block and forces --cache off — the documented
  #250 workaround for the stale-baked-config cache bug — capturing per-variant
  metrics into a comparison table. FORCED_CACHE flips to "rebuild" once #250 lands.

Why

Per the parallel-work plan: #249/#250 block the measurement/gating refinement,
not the tooling that consumes them. Building these now means the post-fix
calibration is a single pass rather than a scramble.

Reuse

• negatives_separation mirrors tools/split_metrics_by_lang.py's NDJSON-read.
• ablate_retrieval reuses run_eval's build_eval_command / parse_eval_report
  / ensure_base / merge_env.

Verification

• 12 new offline tests (both score schemas for #249; pure-helper coverage for #250).
• ruff + mypy src + pytest green (87 passed).
• ablate_retrieval --dry-run builds correct per-variant --cache off commands.

🤖 Generated with Claude Code