exp4 rerun @ 5 seeds × 5 samples + default n_seeds=5

exp8 round 3 showed that single-seed diffusion gives abnormally narrow
sample distributions — on 8t59 the DockQ ≥ 0.23 success rate in the
sample pool went from 44% (1 seed × 25) to 92% (5 seeds × 5). This
likely affects categories beyond ab-ag too.

Changes:
- ensure_bench_run default n_seeds=5 (was 1). Old behavior accessible
  via explicit n_seeds=1. est_wall_hours default bumped to 2h to
  reflect 5× recycling overhead.
- exp4 notebook: new run name `protenix-v1-5seeds`, n_samples=5 +
  n_seeds=5, est_wall_hours=4. Previous single-seed run remains
  cached under `protenix-v1-default` for reference.

Estimated cost ~$126 (over $100 gate; run approved by user as the new
baseline against which all future experiments compare).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: timodonnell

experiments/exp4_baseline_protenix_v1/README.md

@@ -66,20 +66,26 @@ Expected wall clock ~45 min; estimated cost ~$24.
 
 ```python
 bench = ensure_bench_run(
-    "protenix-v1-default",
+    "protenix-v1-5seeds",
     checkpoint="protenix-v1",
     workers=8,
     gpu="H100",
-    n_samples=5,
+    n_samples=5,           # samples per seed
+    n_seeds=5,             # total predictions = 25, matches published FoldBench protocol
     n_cycles=10,
     max_tokens=2048,
     cutoff_date="2024-01-01",
-    est_wall_hours=0.75,
+    est_wall_hours=4.0,    # 5× recycling per target ⇒ ~5× 0.75h
 )
 git_sha = bench.meta.get("git_sha") or "?"
 print(f"cached: {bench.cached}  |  git_sha: {git_sha[:8]}")
 ```
 
+This is the new baseline. The previous single-seed run
+(`protenix-v1-default`) is kept under the old `data/summary.csv` for
+reference; the n=5×5 numbers become the post-exp8-round-3 canonical
+comparison point.
+
 ## Summary table
 
 ```python

src/helico/experiment.py

@@ -304,13 +304,13 @@ def ensure_bench_run(
     workers: int = 8,
     gpu: str = "H100",
     n_samples: int = 5,
-    n_seeds: int = 1,
+    n_seeds: int = 5,
     max_tokens: int = 2048,
     n_cycles: int = 10,
     cutoff_date: str = "2024-01-01",
     categories: str = "",
     target_pdb_ids: str = "",
-    est_wall_hours: float = 0.5,
+    est_wall_hours: float = 2.0,
     force: bool = False,
     publish: bool = False,
 ) -> BenchRun: