test(training): require multi-seed curriculum evaluation

training/generation5.py

@@ -38,6 +38,10 @@ PROMOTED_EASY = REPO_ROOT / "Game" / "bots" / "promoted" / "easy.json"
 MAX_RETRIES = 4
 EVAL_EPISODES = 100
 REGRESSION_MARGIN = 0.15
+# A single paired seed can produce a large physical-side swing even for a
+# policy playing itself. Keep the first historical seed for continuity, but
+# require two independent deterministic sequences before a stage can pass.
+DEFAULT_EVALUATION_SEEDS = (1, 19, 43)
 # --min-head-entropy-frac / --ent-coef-max added 2026-08-24. The aggregate
 # entropy target is a SUM and read healthy (21% of h_max, on target) through
 # all nine Stage-5 attempts while thrust_y alone sat at 14% of its own ceiling
@@ -582,11 +586,12 @@ def run_training(state: dict, stage_index: int, attempt: int, args) -> str:
     return experiment
 
 
-def evaluate(experiment: str, reference: pathlib.Path, args) -> dict:
+def evaluate(experiment: str, reference: pathlib.Path, args, seed: int) -> dict:
     candidate = REPO_ROOT / "Game" / "bots" / f"{experiment}.json"
     cmd = [
         ".venv/bin/python", "evaluate.py", str(candidate), str(reference),
         "--episodes", str(EVAL_EPISODES), "--speedup", str(args.speedup),
+        "--seed", str(seed),
     ]
     if args.godot_bin:
         cmd += ["--godot_bin", args.godot_bin]
@@ -628,11 +633,22 @@ def main() -> None:
     parser.add_argument("--n-parallel", type=int, default=14)
     parser.add_argument("--speedup", type=int, default=16)
     parser.add_argument("--godot-bin", default=None, help="Godot binary for post-stage evaluation")
+    parser.add_argument(
+        "--evaluation-seeds",
+        default=",".join(str(seed) for seed in DEFAULT_EVALUATION_SEEDS),
+        help="Comma-separated independent paired seeds required for every reference evaluation",
+    )
     parser.add_argument("--foundation-checkpoint", default=str(FOUNDATION_CHECKPOINT))
     parser.add_argument("--force-retry", action="store_true")
     parser.add_argument("--skip-to-next-stage", action="store_true")
     parser.add_argument("--dry-run", action="store_true", help="Print the next run command without executing it")
     args = parser.parse_args()
+    try:
+        evaluation_seeds = tuple(dict.fromkeys(int(value) for value in args.evaluation_seeds.split(",") if value.strip()))
+    except ValueError as error:
+        parser.error(f"--evaluation-seeds must be comma-separated integers: {error}")
+    if not evaluation_seeds:
+        parser.error("--evaluation-seeds requires at least one seed")
 
     state = load_state()
     if state["status"] == "done":
@@ -670,7 +686,11 @@ def main() -> None:
         # Preserve order while avoiding a duplicate Stage-5 evaluation in
         # the league stage (its predecessor is also in the pool).
         references = list(dict.fromkeys(references))
-        records = [evaluate(experiment, reference, args) for reference in references]
+        records = [
+            evaluate(experiment, reference, args, seed)
+            for reference in references
+            for seed in evaluation_seeds
+        ]
         match_ok = all(match_passes(record) for record in records)
         evaluation_goal_floor = stage.get("evaluation_goal_rate_floor", 0.0)
         evaluation_goal_failures = [