fix(ml-alpha): captured graph poisoned eval via SyncOnDrop events

z2w9w cluster run hit CUDA_ERROR_INVALID_VALUE at "eval snap_batched
fwd" the first time validation ran after a captured training step.
Training itself succeeded (epoch 0 train_loss=0.69 over 250 captured
graph replays); only the subsequent direct eval kernel launch failed.

Root cause (vendor/cudarc/src/driver/safe/core.rs:920):
  CudaSlice::device_ptr_mut() returns a `SyncOnDrop::Record` guard.
  On drop, that guard UNCONDITIONALLY calls `event.record(stream)` on
  the slice's `.read` event (the check at line 953 only gates the
  cuStreamWaitEvent on .write — the unconditional event.record at the
  end runs no matter what). Inside a stream-capture region, those
  event.record(stream) calls turn the CudaEvents into "captured
  events" per the CUDA Driver API. Captured events can ONLY be waited
  on by streams in the same capture sequence; any later
  cuStreamWaitEvent from outside fails with CUDA_ERROR_INVALID_VALUE.

  The trainer's eval path then called `device_ptr_mut()` again to
  stage the eval DtoDs — which inserted exactly that
  cuStreamWaitEvent on the now-captured `.write` event of every
  trainer CudaSlice. First kernel launch after the dtods failed.

Why this hit z2w9w now: a) all our work is on a SINGLE stream
(`self.stream`), so the event-based multi-stream sync that cudarc
inserts is pure overhead, b) the capture region is exactly where
those overhead events become poisonous.

Fix: disable cudarc's read/write event tracking BEFORE the trainer
allocates ANY device memory. With tracking off at alloc time,
CudaSlice::new returns `read: None, write: None` (core.rs:1283).
SyncOnDrop::record_event with `event: None` produces a `Record(None)`
that does nothing on drop. launch_builder skips its event waits/
records too. The capture region runs clean; the eval direct launches
have no stale captured events to wait on.

Validation:
  - 6 perception_overfit tests pass, including 3 NEW regression tests
    that pin the exact failure modes:
    * evaluate_alone_succeeds — eval with no prior training
    * evaluate_works_after_warmup_only — eval after 1 uncaptured step
    * evaluate_works_after_capture_no_replay — eval right after capture
      (the minimal repro that pinpointed `device_ptr_mut` inside capture)
    * evaluate_works_after_captured_training_step — full warmup +
      capture + replay + eval
  - 26 ml-alpha lib + 23 ml-alpha integration + 306 ml-core lib all pass.

Also drops the now-redundant disable/enable_event_tracking dance
around the capture region — events are globally disabled for the
trainer's lifetime so no per-capture flipping needed.

Honors: feedback_no_quickfixes.md (root-cause traced through
cudarc's safe wrappers to the SyncOnDrop record contract, not a
symptom-suppress sleep/retry hack).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

crates/ml-alpha/src/trainer/perception.rs

@@ -271,6 +271,20 @@ impl PerceptionTrainer {
 
         let stream = dev.cuda_stream().context("trainer stream")?.clone();
         let ctx = dev.cuda_context().context("trainer ctx")?;
+        // Disable cudarc's automatic per-allocation read/write event
+        // tracking BEFORE allocating any device memory. With tracking
+        // enabled, each `device_ptr_mut()` call returns a SyncOnDrop
+        // guard that records a CudaEvent on the stream when dropped.
+        // Inside a stream-capture region those `event.record(stream)`
+        // calls turn the events into "captured events" — which then
+        // cannot be waited on from outside the capture, breaking the
+        // non-captured eval path with CUDA_ERROR_INVALID_VALUE at the
+        // next `cuStreamWaitEvent`. All trainer work runs on a single
+        // stream (`self.stream`) and is therefore stream-ordered with
+        // no cross-stream sync needed, so event tracking is pure
+        // overhead for us. Disabled for the lifetime of this trainer.
+        unsafe { ctx.disable_event_tracking(); }
+        let _ = ctx.check_err();
         let snap_module = ctx.load_cubin(SNAP_CUBIN.to_vec()).context("snap cubin")?;
         let step_module = ctx.load_cubin(STEP_CUBIN.to_vec()).context("step cubin")?;
         let heads_module = ctx.load_cubin(HEADS_CUBIN.to_vec()).context("heads cubin")?;
@@ -599,16 +613,15 @@ impl PerceptionTrainer {
                 .context("train warmup dispatch")?;
             self.cublas_warmed = true;
         } else {
-            let ctx = self.stream.context().clone();
-            let _ = ctx.check_err();
-            unsafe { ctx.disable_event_tracking(); }
-
+            // Event tracking was disabled at trainer construction; the
+            // trainer's CudaSlices have no read/write events, so neither
+            // launch_builder nor device_ptr_mut() will try to insert
+            // cuStreamWaitEvent / event.record() calls inside the
+            // captured region. Capture safely.
             let begin = self.stream.begin_capture(
                 CUstreamCaptureMode::CU_STREAM_CAPTURE_MODE_RELAXED,
             );
             if let Err(e) = begin {
-                unsafe { ctx.enable_event_tracking(); }
-                let _ = ctx.check_err();
                 return Err(anyhow::anyhow!("train begin_capture: {e}"));
             }
 
@@ -617,8 +630,6 @@ impl PerceptionTrainer {
             let graph_result = self.stream.end_capture(
                 CUgraphInstantiate_flags::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
             );
-            unsafe { ctx.enable_event_tracking(); }
-            let _ = ctx.check_err();
 
             dispatch_result.context("train graph dispatch (during capture)")?;
             let graph = graph_result

crates/ml-alpha/tests/perception_overfit.rs

@@ -128,3 +128,118 @@ fn stacked_trainer_loss_shrinks_on_constant_signal() {
          start={initial_avg:.4}, end={final_avg:.4}"
     );
 }
+
+/// Eval alone must work — proves the eval path is fine WITHOUT any
+/// prior captured-graph training. If this passes but
+/// `evaluate_works_after_captured_training_step` fails, the bug is in
+/// the train→eval transition (captured graph leaving state hostile to
+/// direct kernel launches).
+#[test]
+fn evaluate_alone_succeeds() {
+    let dev = test_device();
+    let cfg = PerceptionTrainerConfig {
+        seq_len: 16,
+        mamba2_state_dim: 8,
+        lr_cfc: 3e-3,
+        lr_mamba2: 1e-3,
+        seed: 0x6262,
+        horizon_weights: [1.0; 5],
+        n_batch: 1,
+    };
+    let mut trainer = PerceptionTrainer::new(&dev, &cfg).expect("init");
+    let ts = 1_000_000u64;
+    let prev_mid = 5500.0_f32;
+    let (seq, labels) = synthetic_seq(cfg.seq_len, prev_mid, ts);
+    let (loss, probs) = trainer.evaluate(&seq, labels.as_slice()).expect("eval alone");
+    assert!(loss.is_finite(), "eval loss must be finite, got {loss}");
+    assert_eq!(probs.len(), cfg.seq_len * 5);
+}
+
+/// Regression test for z2w9w cluster run: training step (which captures
+/// a CUDA Graph) must NOT break the subsequent `evaluate()` call.
+/// z2w9w hit CUDA_ERROR_INVALID_VALUE at "eval snap_batched fwd" the
+/// very first time eval ran after training; the captured graph or some
+/// of its launch state left the stream in a state hostile to the
+/// non-captured eval path.
+#[test]
+fn evaluate_works_after_captured_training_step() {
+    let dev = test_device();
+    let cfg = PerceptionTrainerConfig {
+        seq_len: 16,
+        mamba2_state_dim: 8,
+        lr_cfc: 3e-3,
+        lr_mamba2: 1e-3,
+        seed: 0x5151,
+        horizon_weights: [1.0; 5],
+        n_batch: 1,
+    };
+    let mut trainer = PerceptionTrainer::new(&dev, &cfg).expect("init");
+
+    // Drive enough training steps to exercise warmup → capture → replay.
+    let mut ts = 1_000_000u64;
+    let mut prev_mid = 5500.0_f32;
+    for _ in 0..5 {
+        let (seq, labels) = synthetic_seq(cfg.seq_len, prev_mid, ts);
+        trainer.step(&seq, labels.as_slice()).expect("train step");
+        prev_mid = 0.5 * (seq.last().unwrap().bid_px[0] + seq.last().unwrap().ask_px[0]);
+        ts = seq.last().unwrap().ts_ns;
+    }
+
+    // Now call evaluate — must NOT fail with CUDA_ERROR_INVALID_VALUE.
+    let (seq, labels) = synthetic_seq(cfg.seq_len, prev_mid, ts);
+    let (loss, probs) = trainer.evaluate(&seq, labels.as_slice()).expect("evaluate after train");
+    assert!(loss.is_finite(), "eval loss must be finite, got {loss}");
+    assert_eq!(probs.len(), cfg.seq_len * 5, "eval probs must be [K, 5] flat");
+    assert!(probs.iter().all(|p| p.is_finite()), "eval probs must be finite");
+}
+
+/// 2 steps = warmup + capture (no replay yet). Does eval fail right
+/// after capture but BEFORE the first graph.launch?
+#[test]
+fn evaluate_works_after_capture_no_replay() {
+    let dev = test_device();
+    let cfg = PerceptionTrainerConfig {
+        seq_len: 16,
+        mamba2_state_dim: 8,
+        lr_cfc: 3e-3,
+        lr_mamba2: 1e-3,
+        seed: 0x8181,
+        horizon_weights: [1.0; 5],
+        n_batch: 1,
+    };
+    let mut trainer = PerceptionTrainer::new(&dev, &cfg).expect("init");
+    let ts = 1_000_000u64;
+    let prev_mid = 5500.0_f32;
+    let (seq, labels) = synthetic_seq(cfg.seq_len, prev_mid, ts);
+    trainer.step(&seq, labels.as_slice()).expect("warmup step");
+    trainer.step(&seq, labels.as_slice()).expect("capture step");
+    // Eval right after capture, no replay.
+    let (loss, _probs) = trainer.evaluate(&seq, labels.as_slice()).expect("eval after capture");
+    assert!(loss.is_finite(), "eval loss must be finite, got {loss}");
+}
+
+/// Does the bug surface after just ONE step (warmup only, no capture)?
+/// If yes, the warmup dispatch path itself breaks subsequent eval.
+/// If no, the captured graph instantiation / launch is what breaks eval.
+#[test]
+fn evaluate_works_after_warmup_only() {
+    let dev = test_device();
+    let cfg = PerceptionTrainerConfig {
+        seq_len: 16,
+        mamba2_state_dim: 8,
+        lr_cfc: 3e-3,
+        lr_mamba2: 1e-3,
+        seed: 0x7171,
+        horizon_weights: [1.0; 5],
+        n_batch: 1,
+    };
+    let mut trainer = PerceptionTrainer::new(&dev, &cfg).expect("init");
+    let ts = 1_000_000u64;
+    let prev_mid = 5500.0_f32;
+    let (seq, labels) = synthetic_seq(cfg.seq_len, prev_mid, ts);
+    // ONLY one step — warmup, no capture yet.
+    trainer.step(&seq, labels.as_slice()).expect("warmup step");
+    // Now eval.
+    let (loss, _probs) = trainer.evaluate(&seq, labels.as_slice()).expect("eval after warmup");
+    assert!(loss.is_finite(), "eval loss must be finite, got {loss}");
+}