chore(ml-alpha): deep cleanup — delete all V1 dead code
CfcTrunk (~250 lines deleted):
- Deleted V1 forward methods: dispatch_perception, capture_graph_a,
perception_forward_captured, snapshot_hidden, update_input_buffers,
forward_snapshot, upload_pre_allocated
- Deleted V1 weight fields: heads_w_d, heads_b_d, proj_w_d, proj_b_d,
proj_g_d, proj_n_d
- Deleted V1 per-step scratches: h_ping, h_pong, bid_px_d, bid_sz_d,
ask_px_d, ask_sz_d, prev_bid_sz_d, prev_ask_sz_d, regime_d,
snap_feat_d, probs_d, proj_out_d
- Deleted V1 staging buffers: stg_bid_px / stg_bid_sz / stg_ask_px /
stg_ask_sz / stg_regime (MappedF32Buffer was only used by V1)
- Deleted graph_a field + _proj_module + V1 fn handles (snap_fn, step_fn,
heads_fn, proj_fn)
- Deleted V1-only init in new_random (heads_w/b, proj_w/b/g/n, per-step
scratch allocs)
- Deleted file-level helpers used only by V1: upload(stream, host),
upload_into, copy_dtod, download
- Deleted PROJ_CUBIN constant
- Updated save_load_roundtrip test to assert on v2 weight tensors
- Stripped unused imports (CUgraphInstantiate_flags, CUstreamCaptureMode,
CudaGraph, LaunchConfig, PushKernelArg, DevicePtr, DevicePtrMut,
MappedF32Buffer, ES_TICK_SIZE, Mbp10RawInput, REGIME_DIM, PROJ_DIM)
PerceptionTrainer (~30 lines deleted):
- Deleted duplicate cubin fn fields made dead by X10b: snap_batched_fn,
step_batched_fn, heads_grn_fwd_fn, transpose_3d_fn, vsn_fwd_fn,
ln_fwd_fn, attn_fwd_fn (trainer reads these from self.trunk now)
- Deleted their load_function bindings in PerceptionTrainer::new
- Backward kernels (ln_bwd_fn, vsn_bwd_fn, attn_bwd_fn, step_bwd_batched_fn,
heads_grn_bwd_fn) kept — training-only, not on trunk
Verification:
- perception_forward_golden: PASS (max_diff = 0.000000)
- ml-alpha lib tests: 33 pass
- ml-backtesting + fxt-backtest build clean
Trunk.rs shrunk from 800+ to ~480 lines. Code is now purely the v2
inference graph: weights, kernel handles, save_checkpoint/load_checkpoint,
mamba2_l1/l2 accessors. No V1 surface area left.
cecc08a122