sparkring lane: MML A/B negative (collapse identical at 262144) — restore 524288 prod parity; kv usage 0.7% during stall excludes memory

recipes/mimo26-flash-mopd-4x-sparkring/recipe.toml

@@ -63,13 +63,13 @@ args = [
     "--distributed-executor-backend", "mp",
     "--gpu-memory-utilization", "0.85",
     "--block-size", "128",
-    # EXPERIMENT 2026-09-29 (operator): A/B their qualified 262144 against
-    # our 524288 to isolate the c2 long-context collapse axis — pool stays
-    # auto (4 nodes, no memory pressure observed during the stall; their
-    # 20 GiB --kv-cache-memory-bytes pin deliberately not adopted).
-    # NOTE: depth-suite d262144 exceeds 262144 total context → that row
-    # errors until MML goes back up (fail = warn only, suite survives).
-    "--max-model-len", "262144",
+    # EXPERIMENT 2026-09-29 (operator): A/B their qualified 262144 vs our
+    # 524288 on the c2 long-context collapse — RESULT: NEGATIVE, collapse
+    # is bit-identical at both (c2: 13.8 t/s @65K, 5.6 @131K; c1 healthy
+    # 42-48 everywhere) → MML not the axis; back to 524288 for prod parity.
+    # Pool stayed auto throughout (KV usage during the stall: 0.7% — memory
+    # excluded, as the operator predicted).
+    "--max-model-len", "524288",
     "--max-num-seqs", "16",
     "--max-num-batched-tokens", "8192",
     "--model-loader-extra-config", """