feat: citations, eval sidecar/pvalue/trace, nomic prefixes

mcp 2.x already on main. Does not replace the variety chunker
(one chunk per variety is the anti-hallucination contract).

- Numbered [1] citations on search_docs / search_trials
- Eval JSONL sidecar, k-curve, eval.pvalue, eval.trace
- Nomic prefixes at embed time only; stored text unprefixed

Closes #25

eval/run_eval.py

@@ -265,10 +265,15 @@ def main() -> int:
     p = argparse.ArgumentParser()
     p.add_argument("--queries", type=Path, default=Path("eval/queries.jsonl"))
     p.add_argument("--k", type=int, default=5)
+    p.add_argument("--ks", default="1,5,10,20", help="comma-separated k-curve")
     p.add_argument("--output", type=Path, default=Path("eval/results/baseline.md"))
+    p.add_argument("--compare", type=Path, default=None)
+    p.add_argument("--trace", action="store_true")
     p.add_argument("--rerank-url", default=os.environ.get("RERANK_URL", ""))
     p.add_argument("--product-name", default=os.environ.get("PRODUCT_NAME", "crop_seed"))
     args = p.parse_args()
+    ks = sorted({int(x) for x in args.ks.split(",") if x.strip()}) or [args.k]
+    max_k = max(ks + [args.k])
 
     if not args.queries.exists():
         print(f"queries file not found: {args.queries}")
@@ -300,14 +305,40 @@ def main() -> int:
     for r in retrievers:
         print(f"running {r.name}...")
         for q in queries:
-            res = _evaluate_one(r, q, args.k, col)
+            res = _evaluate_one(r, q, max_k, col)
             all_results.append(res)
 
     summary = _aggregate(all_results)
     md = _emit_markdown(queries, all_results, summary, args.k)
+    # k-curve from rank_first_match at max_k
+    md += "\n## k-curve (P@1 / recall from rank_first_match)\n\n"
+    md += "| Retriever | " + " | ".join(f"P@1@k={k}" for k in ks) + " |\n"
+    md += "|" + "---|" * (len(ks) + 1) + "\n"
+    by_r: dict[str, list[dict]] = {}
+    for row in all_results:
+        by_r.setdefault(row["retriever"], []).append(row)
+    for name, rows in by_r.items():
+        cells = []
+        for kk in ks:
+            hits = sum(1 for r in rows if r.get("rank_first_match") and r["rank_first_match"] <= kk)
+            cells.append(f"{hits / len(rows):.3f}" if rows else "0")
+        md += f"| `{name}` | " + " | ".join(cells) + " |\n"
     args.output.parent.mkdir(parents=True, exist_ok=True)
     args.output.write_text(md, encoding="utf-8")
+    sidecar = args.output.with_suffix(".jsonl")
+    with open(sidecar, "w") as fh:
+        for r in all_results:
+            rank = r.get("rank_first_match")
+            fh.write(json.dumps({
+                "query": r["query"],
+                "retriever": r["retriever"],
+                "passed": bool(r.get("passed")),
+                "p_at_1": 1 if rank == 1 else 0,
+                "rr": (1.0 / rank) if rank else 0.0,
+                "rank_first_match": rank,
+            }) + "\n")
     print(f"\nreport: {args.output}")
+    print(f"sidecar: {sidecar}")
     print()
     # Print summary to stdout too
     for line in md.split("\n"):
@@ -315,6 +346,15 @@ def main() -> int:
             print(line)
         if line.startswith("## Per-query"):
             break
+    if args.compare:
+        from eval.pvalue import compare, load_sidecar, render
+        print()
+        print(render(compare(load_sidecar(sidecar), load_sidecar(args.compare))), end="")
+    if args.trace:
+        from eval.trace import render_misses
+        misses_path = args.output.with_name("misses.md")
+        misses_path.write_text(render_misses(all_results))
+        print(f"misses: {misses_path}")
     return 0