plan: amend 4 issues from review — train_step_gpu, mamba2 backward, submit_adam_ops, workspace

A: train_step_gpu must not capture its own graphs (rename to submit_training_ops_ungraphed)
B: Mamba2 backward must be inside unified graph (forward+backward same capture)
C: submit_adam_ops signature verified, needs weight set refs
D: cuBLAS workspace sharing — add verification logging, fail if >32MB needed

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs/superpowers/plans/2026-04-17-unified-cublas-training-graph.md

@@ -1951,6 +1951,109 @@ Task 6 (Attention cubins)──┘
 
 ---
 
+## AMENDMENTS (from plan review)
+
+### Amendment A: train_step_gpu must not capture its own graphs
+
+**Affects:** Task 9 (Step 9.3) and Task 10.
+
+The ungraphed path at step 0 calls `self.trainer.train_step_gpu()` which internally calls `capture_training_graphs()` and creates `graph_forward`, `graph_forward_ddqn`, `graph_adam`. These MUST NOT be created during the unified path — the unified graph captures everything.
+
+**Fix:** In Task 10, when removing `capture_training_graphs()` from `gpu_dqn_trainer.rs`, also modify `train_step_gpu()` to ONLY submit ops (call `submit_forward_ops_main()` + `submit_forward_ops_ddqn()`) WITHOUT capturing any graphs. Rename to `submit_training_ops_ungraphed()` for clarity. The graph capture happens ONLY in `capture_graph_unified()`.
+
+```rust
+/// Submit all training ops ungraphed (step 0 only, before unified graph capture).
+pub(crate) fn submit_training_ops_ungraphed(
+    &mut self,
+    gpu_batch: &GpuBatch,
+    online_d: &DuelingWeightSet,
+    online_b: &BranchingWeightSet,
+    target_d: &DuelingWeightSet,
+    target_b: &BranchingWeightSet,
+) -> Result<(), MLError> {
+    self.upload_batch_gpu(gpu_batch)?;
+    self.update_stochastic_depth_mask()?;
+    self.adam_step += 1;
+    unsafe { *self.t_pinned = self.adam_step; }
+    self.submit_forward_ops_main()?;
+    self.submit_forward_ops_ddqn()?;
+    Ok(())
+}
+```
+
+### Amendment B: Mamba2 backward must be inside unified graph
+
+**Affects:** Task 7 and Task 9.
+
+Mamba2 forward is inside the unified graph (via `submit_forward_ops_main` from Task 7). Mamba2 backward is currently outside the graph (line 1693 in Step 9.3). This means during replay, the backward operates on stale activations from the PREVIOUS capture.
+
+**Fix:** Move `mamba2_backward()` + `step_mamba2_adam()` inside the unified graph capture in Task 9 (after Phase 6 Adam, before Phase 7 ISV). Update `capture_graph_unified()`:
+
+```rust
+    // ── Phase 6: Pruning + grad norm + Adam + unflatten ──
+    let adam_result = if conf_result.is_ok() {
+        self.trainer.apply_pruning_mask()
+            .and_then(|_| self.trainer.submit_adam_ops())
+            .map_err(|e| anyhow::anyhow!("unified adam: {e}"))
+    } else { Ok(()) };
+
+    // ── Phase 6b: Mamba2 BPTT + SGD (must be inside graph with forward) ──
+    let mamba_result = if adam_result.is_ok() {
+        self.mamba2_backward(self.batch_size)
+            .and_then(|_| self.step_mamba2_adam())
+            .map_err(|e| anyhow::anyhow!("unified mamba2 bwd: {e}"))
+    } else { Ok(()) };
+
+    // ── Phase 7: ISV signal update ──
+    let isv_result = if mamba_result.is_ok() {
+        // ...
+```
+
+Remove `self.mamba2_backward()` and `self.step_mamba2_adam()` from the post-graph section of `run_full_step()`.
+
+### Amendment C: submit_adam_ops must exist as public method
+
+**Affects:** Task 9 (Step 9.2).
+
+The capture references `self.trainer.submit_adam_ops()` but the current code's public method is `replay_adam_and_readback()` which replays a pre-captured graph_adam. Task 10 removes graph_adam. The underlying op submission function `submit_adam_ops()` already exists at `gpu_dqn_trainer.rs:8125` — it just needs to be `pub(crate)` (currently it is).
+
+**Verify:** `submit_adam_ops()` signature at gpu_dqn_trainer.rs:8125:
+```rust
+pub(crate) fn submit_adam_ops(
+    &mut self,
+    online_d: &DuelingWeightSet,
+    online_b: &BranchingWeightSet,
+) -> Result<(), MLError>
+```
+
+Note: it takes weight set refs. In the unified capture, these are available as `self.online_dueling` / `self.online_branching`. Update the capture call:
+```rust
+self.trainer.submit_adam_ops(&self.online_dueling, &self.online_branching)
+```
+
+### Amendment D: cuBLAS workspace sharing verification
+
+**Affects:** Tasks 1, 3, 5.
+
+All IQL (8 GEMMs), IQN (~10 GEMMs), and Attention (~12 GEMMs) share the single `lt_workspace_ptr` (32 MB) from `SharedCublasHandle`. During unified graph capture, all GEMMs are stream-ordered (one at a time) so workspace sharing is safe — no concurrent access.
+
+**Verification step:** Add to each GEMM descriptor creation: log the `heuristic.workspaceSize` returned by the heuristic. If ANY GEMM needs >32MB, increase `SharedCublasHandle::lt_workspace_size`.
+
+```rust
+let heuristic = cublaslt_result::get_matmul_algo_heuristic(...)?;
+tracing::info!(%label, ws_needed = heuristic.workspaceSize, "GEMM desc created");
+if heuristic.workspaceSize > lt_ws_size {
+    return Err(MLError::ModelError(format!(
+        "{label}: needs {}MB workspace but only {}MB available",
+        heuristic.workspaceSize / (1024*1024), lt_ws_size / (1024*1024)
+    )));
+}
+```
+
+Add this to Task 1 Step 1.2, Task 3, and Task 5 constructors.
+
+---
+
 ## Task 13: Hardcoded Dimension Cleanup
 
 **Files:**