launcher: adopt D-018 -- no gradient checkpointing, push every 127 steps, plan at the sustained 12,300 tok/s

kernels/phase4_session.py

@@ -45,7 +45,10 @@ CKPT_EVERY = max(1, HORIZON_STEPS // 10)                      # 381 -- §2's one
 
 # Planning rate: the defaults below are P3's measured frozen-geometry figures, so a session launched with
 # no environment set still plans at the rate the gate measured rather than the pre-Gate 3 estimate.
-PLANNING_TOK_PER_S = float(os.environ.get("PLANNING_TOK_PER_S", "9358"))
+# D-018's planning rate: the *sustained* measurement from the two probes (12,221 and 12,312 tok/s over
+# 180 and 120 steps respectively), not the 20-step progress-bar figure of 12,977. Planning on the optimistic
+# number is how a session ends up killed by its own ceiling with an interval still to push.
+PLANNING_TOK_PER_S = float(os.environ.get("PLANNING_TOK_PER_S", "12300"))
 SESSION_GPU_HOURS = float(os.environ.get("SESSION_GPU_HOURS", "6.9"))
 QUOTA_LEFT_HOURS = float(os.environ.get("QUOTA_LEFT_HOURS", "30.0"))
 RESERVE_HOURS = float(os.environ.get("RESERVE_HOURS", "0.6"))  # startup, val eval, the last push
@@ -274,9 +277,20 @@ def main():
     argv = ["train_ounce100m.py",
             "--root", f"{WORK}/mixroot", "--out", f"{WORK}/run",
             "--hub-repo", CKPT_REPO, "--prune",
-            "--seq-len", str(SEQ_LEN), "--attn", "eager", "--grad-ckpt",
+            "--seq-len", str(SEQ_LEN), "--attn", "eager",
+            # D-018, from probe 2 (soak): 180 steps without checkpointing at the frozen geometry held at
+            # 12.84 GiB reserved cross-rank with no drift between step 10 and step 120, and the two probes'
+            # sustained rates (12,221-12,312 tok/s) are +27 % over the checkpointed 9,358. Gradient
+            # checkpointing computes identical gradients, so this changes no frozen hyperparameter -- and
+            # the checkpointed rate does not fit the week's quota at all (D-017 addendum, plan sweep).
+            "--no-grad-ckpt",
             "--micro-batch", str(MICRO_BATCH), "--accum", str(ACCUM),
             "--tokens", str(TOKENS), "--lr", "6e-4", "--warmup-frac", "0.02", "--decay-frac", "0.80",
+            # Every 127 steps: three pushes per 762-step session, and 127 divides 762 so the plan's tiling
+            # and the trainer's whole-multiple guard stay satisfied. D-017-reopened pre-registered the
+            # finer cadence and E-041 killed it at 797 s a push; the bounded read-back made a push cycle
+            # 13-40 s, which buys a worst-case loss of 45 minutes instead of 2.3 hours.
+            "--push-every-steps", "127",
             "--log-every", "20", "--val-tokens", "2000000",
             "--resume", "auto", "--stop-after-steps", str(pl["stop_after_steps"])]