diff --git a/docs/plans/2026-03-07-ml-crate-split-design.md b/docs/plans/2026-03-07-ml-crate-split-design.md
new file mode 100644
index 000000000..1bb8df69a
--- /dev/null
+++ b/docs/plans/2026-03-07-ml-crate-split-design.md
@@ -0,0 +1,231 @@
+# ML Crate Split — Design Document
+
+## Problem
+
+The `ml` crate is 260K lines / 530 files. It compiles as a single `rustc` invocation taking **55.1s** — 69% of the total workspace build. Only 4 of 16 CPU cores are utilized during this phase. Incremental builds after touching any file in `ml/src/` recompile all 260K lines.
+
+## Goal
+
+Split `ml` into sub-crates that compile in parallel, reducing:
+- Clean build time from ~55s to ~35-40s
+- Incremental rebuild after editing a model from ~55s to ~10-15s
+
+## Architecture
+
+```
+ml-core (shared types, traits, utilities, features, data)
+    ├─→ ml-rl (DQN + PPO + cuda_pipeline)          ── parallel ──┐
+    ├─→ ml-supervised (TFT, Mamba, Liquid, etc.)    ── parallel ──┤
+    │                                                              ↓
+    └──────────────────────────→ ml-infra (hyperopt, ensemble, trainers)
+                                        ↓
+                                   ml (thin facade, re-exports)
+```
+
+`ml-rl` and `ml-supervised` compile **fully in parallel** — they have zero dependencies on each other.
+
+## Crate Assignment (Every Module Accounted For)
+
+### ml-core (~50K lines)
+
+Shared infrastructure that 2+ other sub-crates depend on.
+
+**From lib.rs (types defined inline):**
+- `Trade`, `MarketDataSnapshot`, `FeatureVector`, `IntegerTensor`, `UpdateSummary`
+- `HealthStatus`, `MarketRegime`, `CommonTypeError`, `CommonError`
+- `MLError`, `MLResult`, `UnifiedMLResult`
+- `ModelPrediction`, `InferenceResult`, `ModelMetadata`, `Features`, `MLModel` (traits)
+- `PRECISION_FACTOR`, `MAX_INFERENCE_LATENCY_US`
+- `get_training_device()`, `get_training_device_at()`
+- `HFTPerformanceProfile`, `ParallelExecutor`, `LatencyOptimizer`, `ModelRegistry` helpers
+- All `pub use` re-exports (`ErrorCategory`, `Adam`, `ModelType`)
+
+**Modules (moved as-is):**
+- `error.rs` (66 lines)
+- `types/` (49 lines) — `OHLCVBar` etc.
+- `common/` (1,435 lines) — `FactoredAction`, `ExposureLevel`, `CircuitBreaker`
+- `config/` (818 lines)
+- `cuda_compat.rs` (511 lines)
+- `tensor_ops.rs` (143 lines)
+- `traits.rs` (199 lines)
+- `model.rs` (4 lines)
+- `optimizers/` (261 lines) — `Adam`
+- `gradient_accumulation.rs` (274 lines)
+- `gradient_utils.rs` (169 lines)
+- `gpu/` (836 lines) — capabilities, `DeviceConfig`, `memory_profile`
+- `safety/` (6,467 lines) — `MLSafetyManager`, tensor safety
+- `security/` (1,407 lines)
+- `memory_optimization/` (4,576 lines) — quantization (used by TFT)
+- `checkpoint/` (7,867 lines) — `Checkpointable` trait, storage
+- `training/` (1,565 lines) — `UnifiedTrainable`, `CheckpointMetadata`, `TrainingMetrics`
+- `training.rs` (85 lines)
+- `features/` (20,256 lines) — feature extraction, normalization, OFI, MBP-10
+- `feature_cache.rs` (349 lines)
+- `preprocessing.rs` (754 lines)
+- `data_loader.rs` (713 lines)
+- `data_loaders/` (4,365 lines) — `DbnSequenceLoader`, calibration
+- `data_pipeline/` (1,512 lines)
+- `data_validation/` (3,529 lines)
+- `evaluation/` (1,011 lines) — `EvaluationEngine`, `PerformanceMetrics`
+- `metrics/` (420 lines)
+- `performance.rs` (481 lines)
+- `observability/` (1,193 lines)
+
+**Extracted from dqn/ → ml-core shared modules:**
+- `dqn/mixed_precision.rs` (722 lines) → `ml-core/src/mixed_precision.rs`
+- `dqn/xavier_init.rs` (200 lines) → `ml-core/src/xavier_init.rs`
+- `dqn/portfolio_tracker.rs` (931 lines) → `ml-core/src/portfolio_tracker.rs`
+- `dqn/action_space.rs` (245 lines) → `ml-core/src/action_space.rs`
+- `dqn/order_router.rs` (175 lines) → `ml-core/src/order_router.rs`
+
+**NOT extracting:**
+- `dqn/circuit_breaker.rs` — already a re-export of `common/circuit_breaker.rs` (which stays in ml-core)
+- `liquid/FixedPoint` — only used by mamba (both in ml-supervised), no extraction needed
+
+### ml-rl (~44K lines)
+
+Reinforcement learning models. Depends on: `ml-core`.
+
+- `dqn/` (33,116 lines minus 2,273 extracted = ~30,843 lines)
+- `ppo/` (14,182 lines)
+- `cuda_pipeline/` (4,431 lines) — GPU data pipeline for DQN/PPO trainers
+
+**Import changes:**
+- `crate::dqn::mixed_precision::training_dtype` → `ml_core::mixed_precision::training_dtype`
+- `crate::dqn::xavier_init::linear_xavier` → `ml_core::xavier_init::linear_xavier`
+- `crate::dqn::portfolio_tracker` → `ml_core::portfolio_tracker`
+- `crate::dqn::action_space` → `ml_core::action_space`
+- `crate::dqn::order_router` → `ml_core::order_router`
+- `crate::dqn::circuit_breaker` → `ml_core::common::circuit_breaker`
+- All `crate::MLError` → `ml_core::MLError`
+- All `crate::training::unified_trainer` → `ml_core::training::unified_trainer`
+
+### ml-supervised (~36K lines)
+
+Supervised/generative models. Depends on: `ml-core`. **Zero dependency on ml-rl.**
+
+- `tft/` (11,440 lines)
+- `mamba/` (6,465 lines)
+- `liquid/` (5,228 lines) — includes `FixedPoint`
+- `tgnn/` (4,410 lines)
+- `tlob/` (2,189 lines)
+- `kan/` (1,397 lines)
+- `xlstm/` (1,404 lines)
+- `diffusion/` (1,317 lines)
+- `transformers/` (821 lines) — shared transformer utilities
+- `flash_attention/` (1,225 lines) — shared attention implementation
+
+**Import changes:**
+- Same pattern as ml-rl: `crate::X` → `ml_core::X` for shared items
+- `crate::liquid::FixedPoint` stays as `crate::liquid::FixedPoint` (both in same crate)
+- `crate::dqn::mixed_precision::training_dtype` → `ml_core::mixed_precision::training_dtype`
+
+### ml-infra (~67K lines)
+
+Training infrastructure. Depends on: `ml-core`, `ml-rl`, `ml-supervised`.
+
+- `hyperopt/` (19,934 lines) — Bayesian optimization, adapters for all 10 models
+- `ensemble/` (13,744 lines) — inference adapters, confidence, voting
+- `trainers/` (17,772 lines) — `DQNTrainer`, `PPOTrainer`, unified training loop
+- `benchmark/` (5,969 lines) — model benchmarks
+- `deployment/` (8,190 lines) — A/B testing, hot-swap, versioning
+- `training_pipeline.rs` (847 lines)
+- `walk_forward.rs` (485 lines)
+
+**Import changes:**
+- `crate::dqn::DQN` → `ml_rl::dqn::DQN`
+- `crate::ppo::PPO` → `ml_rl::ppo::PPO`
+- `crate::tft::TemporalFusionTransformer` → `ml_supervised::tft::TemporalFusionTransformer`
+- (and so on for all model types)
+- Shared items: `crate::X` → `ml_core::X`
+
+### ml (thin facade, ~30K lines)
+
+Re-exports everything from sub-crates. Contains modules that cross sub-crate boundaries or are only used externally.
+
+**Modules (moved here):**
+- `inference.rs` (1,677 lines) — depends on TFT (supervised) + safety (core)
+- `inference_validator.rs` (528 lines)
+- `model_factory.rs` (339 lines) — creates any model type
+- `model_loader_integration.rs` (725 lines)
+- `model_registry/` (558 lines) + `model_registry.rs` (1,350 lines)
+- `registry/` (527 lines)
+- `bridge.rs` (356 lines)
+- `batch_processing.rs` (710 lines)
+- `benchmarks.rs` (733 lines)
+- `examples.rs` (892 lines)
+- `portfolio_transformer.rs` (814 lines)
+- `validation/` (4,337 lines) — depends on DQN + PPO types
+- `risk/` (5,023 lines) — ML risk (Kelly, position sizing)
+- `regime/` (4,943 lines) + `regime_detection/` (1,196 lines)
+- `microstructure/` (2,598 lines)
+- `labeling/` (3,340 lines)
+- `universe/` (1,731 lines)
+- `asset_selection/` (1,184 lines)
+- `backtesting/` (1,120 lines)
+- `paper_trading/` (389 lines)
+- `stress_testing/` (1,254 lines)
+- `integration/` (3,678 lines)
+- `tests/` (1,844 lines)
+- `gpu_benchmarks/` (21 lines)
+- `cuda_common/` (0 lines — empty)
+
+**Re-exports (preserves external API):**
+```rust
+// ml/src/lib.rs
+pub use ml_core::*;          // All types, errors, traits
+pub use ml_rl::dqn;          // DQN module
+pub use ml_rl::ppo;          // PPO module
+pub use ml_rl::cuda_pipeline; // GPU pipeline
+pub use ml_supervised::tft;   // All supervised models
+pub use ml_supervised::mamba;
+// ... etc for all model modules
+pub use ml_infra::hyperopt;
+pub use ml_infra::ensemble;
+pub use ml_infra::trainers;
+// ... etc
+```
+
+## Feature Flags
+
+| Feature | Crate | Reason |
+|---------|-------|--------|
+| `cuda` | `ml-core` (propagated) | candle-core/cuda, candle-nn/cuda |
+| `minimal-inference` | `ml` facade | controls which sub-crates are included |
+| `financial` | `ml` facade | feature gate for services |
+| `s3-storage` | `ml-core` | AWS SDK deps |
+| `high-precision` | `ml-core` | rust_decimal feature |
+| `mimalloc-allocator` | `ml-core` | allocator |
+| `simd` | `ml-core` | SIMD features |
+| `gc` | `ml-core` | GC features |
+| `nccl` | `ml-rl` | multi-GPU for RL training |
+
+## External Consumer Migration
+
+All external consumers (`trading_service`, `ml_training_service`, etc.) continue to depend on `ml` (the facade). The facade re-exports everything, so **zero import changes** in downstream crates. The workspace `Cargo.toml` entry stays: `ml = { path = "crates/ml" }`.
+
+## Projected Compile Times
+
+| Phase | Time | Parallelism |
+|-------|------|-------------|
+| ml-core | ~15s | serial (first) |
+| ml-rl + ml-supervised | ~13s | **2-way parallel** |
+| ml-infra | ~18s | serial (after models) |
+| ml facade | ~3s | serial (last) |
+| **Total clean** | **~49s** | vs 55s current |
+| **Edit in dqn (incremental)** | **~13s** | vs 55s current |
+| **Edit in tft (incremental)** | **~10s** | vs 55s current |
+| **Edit in features (incremental)** | **~10s** | vs 55s current |
+
+## Risks
+
+1. **Circular dependencies**: `validation/` imports both DQN and PPO types → placed in facade
+2. **Test placement**: Tests that cross model boundaries stay in the facade crate
+3. **Feature flag propagation**: `cuda` must be forwarded from facade through sub-crates
+4. **Re-export completeness**: Must verify every `use ml::X` in the workspace still resolves
+
+## Non-Goals
+
+- No stubs — every function is wired to its real implementation
+- No new functionality — pure mechanical restructuring
+- No API changes for external consumers
diff --git a/docs/plans/2026-03-07-ml-crate-split-plan.md b/docs/plans/2026-03-07-ml-crate-split-plan.md
new file mode 100644
index 000000000..096592d71
--- /dev/null
+++ b/docs/plans/2026-03-07-ml-crate-split-plan.md
@@ -0,0 +1,910 @@
+# ML Crate Split Implementation Plan
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Split the monolithic `ml` crate (260K lines, 55s compile) into 5 crates that compile in parallel, reducing incremental rebuilds from 55s to ~10-15s.
+
+**Architecture:** 4 sub-crates (`ml-core`, `ml-rl`, `ml-supervised`, `ml-infra`) + thin `ml` facade that re-exports everything. External consumers see no API changes.
+
+**Tech Stack:** Rust workspace, Cargo features, `pub use` re-exports.
+
+**Reference:** See `docs/plans/2026-03-07-ml-crate-split-design.md` for full module inventory and dependency analysis.
+
+---
+
+## Phase 1: Preparation (in worktree `.worktrees/ml-crate-split`, branch `feature/ml-crate-split`)
+
+### Task 1: Verify Clean Baseline
+
+**Files:** None modified
+
+**Step 1: Run full workspace check**
+
+Run: `cd /home/jgrusewski/Work/foxhunt/.worktrees/ml-crate-split && SQLX_OFFLINE=true cargo check --workspace 2>&1 | tail -5`
+Expected: `Finished` with 0 errors
+
+**Step 2: Run ml tests**
+
+Run: `SQLX_OFFLINE=true cargo test -p ml --lib 2>&1 | tail -5`
+Expected: `test result: ok` with 2506+ tests
+
+**Step 3: Record baseline**
+
+Note the exact test count and build time for comparison after split.
+
+---
+
+### Task 2: Create Sub-Crate Directory Structure
+
+**Files:**
+- Create: `crates/ml-core/Cargo.toml`
+- Create: `crates/ml-core/src/lib.rs`
+- Create: `crates/ml-rl/Cargo.toml`
+- Create: `crates/ml-rl/src/lib.rs`
+- Create: `crates/ml-supervised/Cargo.toml`
+- Create: `crates/ml-supervised/src/lib.rs`
+- Create: `crates/ml-infra/Cargo.toml`
+- Create: `crates/ml-infra/src/lib.rs`
+- Modify: `Cargo.toml` (workspace root — add members + workspace deps)
+
+**Step 1: Create crate directories**
+
+```bash
+mkdir -p crates/ml-core/src crates/ml-rl/src crates/ml-supervised/src crates/ml-infra/src
+```
+
+**Step 2: Add workspace members**
+
+In root `Cargo.toml`, add to `[workspace] members`:
+```toml
+"crates/ml-core",
+"crates/ml-rl",
+"crates/ml-supervised",
+"crates/ml-infra",
+```
+
+And add workspace dependency entries:
+```toml
+ml-core = { path = "crates/ml-core" }
+ml-rl = { path = "crates/ml-rl" }
+ml-supervised = { path = "crates/ml-supervised" }
+ml-infra = { path = "crates/ml-infra" }
+```
+
+**Step 3: Create minimal Cargo.toml for each sub-crate**
+
+Each sub-crate gets a `Cargo.toml` with:
+- Same edition/rust-version as `ml`
+- Dependencies subset (only what its modules need)
+- Feature flags relevant to its modules
+- `ml-core` as dependency for `ml-rl`, `ml-supervised`, `ml-infra`
+
+See the detailed Cargo.toml specs in Tasks 5-8.
+
+**Step 4: Create stub lib.rs for each**
+
+Each gets an empty `pub mod` list initially:
+```rust
+// crates/ml-core/src/lib.rs
+// Placeholder — modules moved in subsequent tasks
+```
+
+**Step 5: Verify workspace compiles**
+
+Run: `SQLX_OFFLINE=true cargo check --workspace 2>&1 | tail -5`
+Expected: Compiles (empty crates are valid)
+
+**Step 6: Commit**
+
+```bash
+git add crates/ml-core crates/ml-rl crates/ml-supervised crates/ml-infra Cargo.toml
+git commit -m "feat(ml): scaffold sub-crate directory structure for ml split"
+```
+
+---
+
+## Phase 2: Build ml-core (Foundation)
+
+### Task 3: Extract Shared Items from dqn/ to Top-Level Modules
+
+These items live in `dqn/` but are used by PPO, trainers, ensemble, hyperopt — they belong in `ml-core`.
+
+**Files:**
+- Move: `crates/ml/src/dqn/mixed_precision.rs` → `crates/ml/src/mixed_precision.rs`
+- Move: `crates/ml/src/dqn/xavier_init.rs` → `crates/ml/src/xavier_init.rs`
+- Move: `crates/ml/src/dqn/portfolio_tracker.rs` → `crates/ml/src/portfolio_tracker.rs`
+- Move: `crates/ml/src/dqn/action_space.rs` → `crates/ml/src/action_space.rs`
+- Move: `crates/ml/src/dqn/order_router.rs` → `crates/ml/src/order_router.rs`
+- Modify: `crates/ml/src/dqn/mod.rs` — replace module declarations with re-exports
+- Modify: `crates/ml/src/lib.rs` — add `pub mod` for new top-level modules
+- Modify: All files importing these via `crate::dqn::*` path
+
+**Step 1: Move files**
+
+```bash
+cd crates/ml/src
+mv dqn/mixed_precision.rs mixed_precision.rs
+mv dqn/xavier_init.rs xavier_init.rs
+mv dqn/portfolio_tracker.rs portfolio_tracker.rs
+mv dqn/action_space.rs action_space.rs
+mv dqn/order_router.rs order_router.rs
+```
+
+**Step 2: Update dqn/mod.rs**
+
+Replace the old `pub mod` declarations with re-exports so existing `crate::dqn::X` paths still work:
+```rust
+// In dqn/mod.rs, REPLACE:
+//   pub mod mixed_precision;
+//   pub mod xavier_init;
+//   pub mod portfolio_tracker;
+//   pub mod action_space;
+//   pub mod order_router;
+// WITH:
+pub use crate::mixed_precision;
+pub use crate::xavier_init;
+pub use crate::portfolio_tracker;
+pub use crate::action_space;
+pub use crate::order_router;
+```
+
+**Step 3: Add modules to ml/lib.rs**
+
+Add these lines to `lib.rs`:
+```rust
+pub mod mixed_precision;
+pub mod xavier_init;
+pub mod portfolio_tracker;
+pub mod action_space;
+pub mod order_router;
+```
+
+**Step 4: Fix internal imports in moved files**
+
+In each moved file, change `use crate::dqn::` imports to `use crate::` where referencing peer modules (e.g., `mixed_precision.rs` may import from `dqn` internals — fix those paths).
+
+Specific fixes needed:
+- `action_space.rs`: has no `crate::dqn::` imports (only std/serde) — no changes needed
+- `xavier_init.rs`: uses `candle_core`/`candle_nn` only — no changes needed
+- `mixed_precision.rs`: uses `candle_core` only — no changes needed
+- `portfolio_tracker.rs`: check for `crate::dqn::` imports and fix
+- `order_router.rs`: imports `crate::dqn::action_space::*` → change to `crate::action_space::*`; imports `crate::common::action::*` → keep as-is
+
+**Step 5: Verify compilation**
+
+Run: `SQLX_OFFLINE=true cargo check -p ml 2>&1 | tail -5`
+Expected: Compiles with 0 errors (re-exports preserve all paths)
+
+**Step 6: Run tests**
+
+Run: `SQLX_OFFLINE=true cargo test -p ml --lib 2>&1 | tail -5`
+Expected: Same test count as baseline, 0 failures
+
+**Step 7: Run clippy**
+
+Run: `SQLX_OFFLINE=true cargo clippy -p ml -- -D warnings 2>&1 | tail -10`
+Expected: 0 errors, 0 warnings
+
+**Step 8: Commit**
+
+```bash
+git add -A crates/ml/
+git commit -m "refactor(ml): extract shared items from dqn/ to top-level modules
+
+Move mixed_precision, xavier_init, portfolio_tracker, action_space,
+order_router from dqn/ to ml/src/ root. These are shared infrastructure
+used by PPO, trainers, hyperopt, and ensemble — not DQN-specific.
+Re-export from dqn/mod.rs for backward compatibility."
+```
+
+---
+
+### Task 4: Write ml-core Cargo.toml
+
+**Files:**
+- Create: `crates/ml-core/Cargo.toml`
+
+**Step 1: Write Cargo.toml**
+
+Build by examining `crates/ml/Cargo.toml` and including ONLY dependencies used by ml-core modules. Key deps:
+- `candle-core`, `candle-nn`, `candle-optimisers` (for mixed_precision, xavier_init, Adam, gpu)
+- `common`, `config`, `data` (workspace crates)
+- `tokio`, `serde`, `thiserror`, `anyhow`, `chrono`, `tracing` (core utilities)
+- `arrow`, `parquet` (for features/)
+- `dbn`, `databento` (for data_loaders/)
+- `rust_decimal` (financial types)
+- `nalgebra`, `ndarray`, `num-traits` (for features math)
+- `dashmap`, `lru` (for caching in features/)
+- `sha2`, `hmac`, `hex` (for security/)
+- `flate2`, `memmap2` (for checkpoint/)
+- NOT: `argmin` (hyperopt only), NOT model-specific deps
+
+Feature flags:
+```toml
+[features]
+default = ["cuda"]
+cuda = ["candle-core/cuda", "candle-core/cudnn", "candle-nn/cuda", "candle-nn/cudnn"]
+s3-storage = ["aws-config", "aws-sdk-s3", "aws-types", "aws-credential-types", "urlencoding"]
+high-precision = ["rust_decimal/serde-float"]
+mimalloc-allocator = ["mimalloc"]
+simd = []
+gc = []
+financial = []
+minimal-inference = []
+```
+
+**Step 2: Verify it parses**
+
+Run: `cargo check -p ml-core 2>&1 | head -5` (will fail on missing modules, that's OK — just verifying Cargo.toml syntax)
+
+**Step 3: Commit**
+
+```bash
+git add crates/ml-core/Cargo.toml
+git commit -m "feat(ml-core): write Cargo.toml with dependency subset"
+```
+
+---
+
+### Task 5: Move Core Modules to ml-core
+
+This is the largest single task. Move all modules assigned to ml-core.
+
+**Files:**
+- Move: ~30 modules from `crates/ml/src/` to `crates/ml-core/src/`
+- Modify: `crates/ml-core/src/lib.rs` — declare all modules
+- Modify: `crates/ml/src/lib.rs` — remove moved modules, add `pub use ml_core::*`
+
+**Step 1: Move type definitions from lib.rs**
+
+Extract the type definitions (`Trade`, `MarketRegime`, `CommonError`, `CommonTypeError`, `MLError`, `HealthStatus`, `MarketDataSnapshot`, `FeatureVector`, `IntegerTensor`, `UpdateSummary`, `ModelPrediction`, `InferenceResult`, `ModelMetadata`, `Features`, `MLModel`, `HFTPerformanceProfile`, `ParallelExecutor`, `LatencyOptimizer`, `ModelRegistry` helpers, and all `From` impls for `MLError`) from `crates/ml/src/lib.rs` into `crates/ml-core/src/lib.rs`.
+
+Keep the crate-level attributes (`#![deny(...)]`, `#![allow(...)]`) in both — ml-core gets the canonical set, ml facade gets a minimal subset.
+
+**Step 2: Move modules**
+
+```bash
+# Core infrastructure
+mv crates/ml/src/error.rs crates/ml-core/src/
+mv crates/ml/src/types crates/ml-core/src/
+mv crates/ml/src/common crates/ml-core/src/
+mv crates/ml/src/config crates/ml-core/src/
+mv crates/ml/src/cuda_compat.rs crates/ml-core/src/
+mv crates/ml/src/tensor_ops.rs crates/ml-core/src/
+mv crates/ml/src/traits.rs crates/ml-core/src/
+mv crates/ml/src/model.rs crates/ml-core/src/
+mv crates/ml/src/optimizers crates/ml-core/src/
+mv crates/ml/src/gradient_accumulation.rs crates/ml-core/src/
+mv crates/ml/src/gradient_utils.rs crates/ml-core/src/
+mv crates/ml/src/gpu crates/ml-core/src/
+mv crates/ml/src/safety crates/ml-core/src/
+mv crates/ml/src/security crates/ml-core/src/
+mv crates/ml/src/memory_optimization crates/ml-core/src/
+mv crates/ml/src/checkpoint crates/ml-core/src/
+mv crates/ml/src/training crates/ml-core/src/
+mv crates/ml/src/training.rs crates/ml-core/src/
+
+# Extracted shared items
+mv crates/ml/src/mixed_precision.rs crates/ml-core/src/
+mv crates/ml/src/xavier_init.rs crates/ml-core/src/
+mv crates/ml/src/portfolio_tracker.rs crates/ml-core/src/
+mv crates/ml/src/action_space.rs crates/ml-core/src/
+mv crates/ml/src/order_router.rs crates/ml-core/src/
+
+# Features and data
+mv crates/ml/src/features crates/ml-core/src/
+mv crates/ml/src/feature_cache.rs crates/ml-core/src/
+mv crates/ml/src/preprocessing.rs crates/ml-core/src/
+mv crates/ml/src/data_loader.rs crates/ml-core/src/
+mv crates/ml/src/data_loaders crates/ml-core/src/
+mv crates/ml/src/data_pipeline crates/ml-core/src/
+mv crates/ml/src/data_validation crates/ml-core/src/
+mv crates/ml/src/evaluation crates/ml-core/src/
+
+# Metrics and monitoring
+mv crates/ml/src/metrics crates/ml-core/src/
+mv crates/ml/src/performance.rs crates/ml-core/src/
+mv crates/ml/src/observability crates/ml-core/src/
+```
+
+**Step 3: Update ml-core/src/lib.rs**
+
+Declare all moved modules with `pub mod`. Include all type definitions and re-exports. Copy the relevant `use` imports (serde, chrono, rust_decimal, etc.).
+
+**Step 4: Update all `use crate::` imports in moved files**
+
+Every file in `crates/ml-core/src/` that uses `use crate::` — these references are now correct (they refer to ml-core's own modules). Verify no moved file imports something that stayed in `ml`.
+
+Potential issues:
+- `safety/` imports from `crate::dqn::` (but mixed_precision is now in ml-core root)
+- `features/` may import from model-specific modules → fix or feature-gate
+- `data_loaders/` imports `crate::types::OHLCVBar` → still in ml-core ✓
+- `evaluation/` imports `crate::features::` → still in ml-core ✓
+- `training/unified_trainer.rs` — verify it only imports items now in ml-core
+- `checkpoint/` — verify no model-specific imports
+
+Run `grep -rn 'use crate::dqn\|use crate::ppo\|use crate::tft\|use crate::mamba\|use crate::ensemble\|use crate::hyperopt\|use crate::trainers' crates/ml-core/src/` to find any imports of modules that are NOT in ml-core. These must be removed/refactored.
+
+**Step 5: Update ml/src/lib.rs**
+
+Replace removed `pub mod` declarations with re-exports:
+```rust
+// Add dependency
+pub use ml_core;
+
+// Re-export everything from ml-core
+pub use ml_core::*;
+// Explicitly re-export modules for path compatibility
+pub use ml_core::common;
+pub use ml_core::config;
+pub use ml_core::checkpoint;
+// ... etc for all modules that external consumers access by path (ml::checkpoint::X)
+```
+
+**Step 6: Add ml-core dependency to ml/Cargo.toml**
+
+```toml
+[dependencies]
+ml-core = { workspace = true }
+```
+
+**Step 7: Verify compilation**
+
+Run: `SQLX_OFFLINE=true cargo check -p ml-core 2>&1 | tail -5`
+Expected: ml-core compiles
+
+Run: `SQLX_OFFLINE=true cargo check -p ml 2>&1 | tail -5`
+Expected: ml compiles (re-exports resolve)
+
+**Step 8: Fix any compilation errors**
+
+Iterate on import fixes until both `ml-core` and `ml` compile cleanly.
+
+**Step 9: Run tests**
+
+Run: `SQLX_OFFLINE=true cargo test -p ml --lib 2>&1 | tail -5`
+Expected: All tests still pass
+
+**Step 10: Run clippy**
+
+Run: `SQLX_OFFLINE=true cargo clippy -p ml-core -p ml -- -D warnings 2>&1 | tail -10`
+Expected: 0 errors, 0 warnings
+
+**Step 11: Commit**
+
+```bash
+git add -A
+git commit -m "refactor(ml): move core modules to ml-core crate
+
+Move types, errors, traits, features, data loaders, checkpoint,
+safety, GPU utils, and shared infrastructure to ml-core.
+Re-export from ml facade for backward compatibility."
+```
+
+---
+
+## Phase 3: Build ml-rl (RL Models)
+
+### Task 6: Write ml-rl Cargo.toml
+
+**Files:**
+- Create: `crates/ml-rl/Cargo.toml`
+
+**Step 1: Write Cargo.toml**
+
+Dependencies: `ml-core` + `candle-core`, `candle-nn` + RL-specific deps (`crossbeam`, etc.).
+
+Feature flags:
+```toml
+[features]
+default = ["cuda"]
+cuda = ["ml-core/cuda"]
+nccl = ["cuda"]
+```
+
+**Step 2: Commit**
+
+---
+
+### Task 7: Move RL Modules to ml-rl
+
+**Files:**
+- Move: `crates/ml/src/dqn/` → `crates/ml-rl/src/dqn/`
+- Move: `crates/ml/src/ppo/` → `crates/ml-rl/src/ppo/`
+- Move: `crates/ml/src/cuda_pipeline/` → `crates/ml-rl/src/cuda_pipeline/`
+- Modify: `crates/ml-rl/src/lib.rs`
+- Modify: `crates/ml/src/lib.rs`
+
+**Step 1: Move modules**
+
+```bash
+mv crates/ml/src/dqn crates/ml-rl/src/
+mv crates/ml/src/ppo crates/ml-rl/src/
+mv crates/ml/src/cuda_pipeline crates/ml-rl/src/
+```
+
+**Step 2: Write ml-rl/src/lib.rs**
+
+```rust
+// Same clippy attributes as ml-core
+pub mod dqn;
+pub mod ppo;
+pub mod cuda_pipeline;
+```
+
+**Step 3: Fix all imports in ml-rl**
+
+In every file under `crates/ml-rl/src/`:
+- `use crate::MLError` → `use ml_core::MLError`
+- `use crate::mixed_precision` → `use ml_core::mixed_precision`
+- `use crate::xavier_init` → `use ml_core::xavier_init`
+- `use crate::portfolio_tracker` → `use ml_core::portfolio_tracker`
+- `use crate::action_space` → `use ml_core::action_space`
+- `use crate::order_router` → `use ml_core::order_router`
+- `use crate::common::` → `use ml_core::common::`
+- `use crate::cuda_compat::` → `use ml_core::cuda_compat::`
+- `use crate::gradient_accumulation::` → `use ml_core::gradient_accumulation::`
+- `use crate::training::unified_trainer::` → `use ml_core::training::unified_trainer::`
+- `use crate::safety::` → `use ml_core::safety::`
+- `use crate::tensor_ops::` → `use ml_core::tensor_ops::`
+- `use crate::gpu::` → `use ml_core::gpu::`
+- `use crate::checkpoint::` → `use ml_core::checkpoint::`
+- `use crate::features::` → `use ml_core::features::`
+- `use crate::evaluation::` → `use ml_core::evaluation::`
+- `use crate::Adam` → `use ml_core::Adam`
+- `use crate::PRECISION_FACTOR` → `use ml_core::PRECISION_FACTOR`
+- `use crate::{MLError, MLResult}` → `use ml_core::{MLError, MLResult}`
+
+Keep `use crate::dqn::` and `use crate::ppo::` as-is (these reference ml-rl's own modules).
+
+Remove the re-exports from `dqn/mod.rs` added in Task 3 (they re-exported from `crate::` which was ml, now those modules are in ml-core):
+```rust
+// REMOVE these from dqn/mod.rs:
+// pub use crate::mixed_precision;
+// pub use crate::xavier_init;
+// etc.
+// REPLACE WITH:
+pub use ml_core::mixed_precision;
+pub use ml_core::xavier_init;
+pub use ml_core::portfolio_tracker;
+pub use ml_core::action_space;
+pub use ml_core::order_router;
+```
+
+Also fix PPO re-exports in `ppo/mod.rs`:
+```rust
+// ppo/circuit_breaker.rs re-exports from crate::common::circuit_breaker
+// Change to: ml_core::common::circuit_breaker
+```
+
+Handle cross-RL deps:
+- PPO imports `crate::dqn::circuit_breaker` → becomes `crate::dqn::circuit_breaker` (still works via dqn re-export) OR `ml_core::common::circuit_breaker` (direct)
+- PPO imports `crate::dqn::mixed_precision::training_dtype` → `ml_core::mixed_precision::training_dtype`
+- PPO imports `crate::dqn::portfolio_tracker` → `ml_core::portfolio_tracker`
+- PPO imports `crate::dqn::xavier_init::linear_xavier` → `ml_core::xavier_init::linear_xavier`
+- cuda_pipeline imports `crate::dqn::mixed_precision` → `ml_core::mixed_precision`
+- cuda_pipeline imports `crate::dqn::replay_buffer_type::GpuBatch` → `crate::dqn::replay_buffer_type::GpuBatch` (stays in ml-rl)
+
+Watch for: `use crate::trainers::DQNTrainer` in dqn/ — this is a reverse dep. If it exists, it's likely in test code. If in non-test code, it needs to be removed/refactored (trainers will be in ml-infra).
+
+Watch for: `use crate::hyperopt::adapters::PPOTrainer` in ppo/ — same issue. Check if test-only.
+
+**Step 4: Update ml/src/lib.rs**
+
+Add re-exports:
+```rust
+pub use ml_rl::dqn;
+pub use ml_rl::ppo;
+pub use ml_rl::cuda_pipeline;
+```
+
+Add dependency to `ml/Cargo.toml`:
+```toml
+ml-rl = { workspace = true }
+```
+
+**Step 5: Verify compilation**
+
+Run: `SQLX_OFFLINE=true cargo check -p ml-rl 2>&1 | tail -10`
+Run: `SQLX_OFFLINE=true cargo check -p ml 2>&1 | tail -10`
+
+**Step 6: Fix errors iteratively**
+
+Keep fixing imports until both compile.
+
+**Step 7: Run tests**
+
+Run: `SQLX_OFFLINE=true cargo test -p ml --lib 2>&1 | tail -5`
+
+**Step 8: Commit**
+
+```bash
+git add -A
+git commit -m "refactor(ml): move DQN, PPO, cuda_pipeline to ml-rl crate"
+```
+
+---
+
+## Phase 4: Build ml-supervised (Supervised Models)
+
+### Task 8: Write ml-supervised Cargo.toml
+
+**Files:**
+- Create: `crates/ml-supervised/Cargo.toml`
+
+Dependencies: `ml-core` + `candle-core`, `candle-nn`.
+
+Feature flags:
+```toml
+[features]
+default = ["cuda"]
+cuda = ["ml-core/cuda"]
+```
+
+**Step 1: Commit**
+
+---
+
+### Task 9: Move Supervised Models to ml-supervised
+
+**Files:**
+- Move: `crates/ml/src/tft/` → `crates/ml-supervised/src/tft/`
+- Move: `crates/ml/src/mamba/` → `crates/ml-supervised/src/mamba/`
+- Move: `crates/ml/src/liquid/` → `crates/ml-supervised/src/liquid/`
+- Move: `crates/ml/src/tgnn/` → `crates/ml-supervised/src/tgnn/`
+- Move: `crates/ml/src/tlob/` → `crates/ml-supervised/src/tlob/`
+- Move: `crates/ml/src/kan/` → `crates/ml-supervised/src/kan/`
+- Move: `crates/ml/src/xlstm/` → `crates/ml-supervised/src/xlstm/`
+- Move: `crates/ml/src/diffusion/` → `crates/ml-supervised/src/diffusion/`
+- Move: `crates/ml/src/transformers/` → `crates/ml-supervised/src/transformers/`
+- Move: `crates/ml/src/flash_attention/` → `crates/ml-supervised/src/flash_attention/`
+- Modify: `crates/ml-supervised/src/lib.rs`
+- Modify: `crates/ml/src/lib.rs`
+
+**Step 1: Move modules**
+
+```bash
+for mod in tft mamba liquid tgnn tlob kan xlstm diffusion transformers flash_attention; do
+  mv crates/ml/src/$mod crates/ml-supervised/src/
+done
+```
+
+**Step 2: Write ml-supervised/src/lib.rs**
+
+```rust
+pub mod tft;
+pub mod mamba;
+pub mod liquid;
+pub mod tgnn;
+pub mod tlob;
+pub mod kan;
+pub mod xlstm;
+pub mod diffusion;
+pub mod transformers;
+pub mod flash_attention;
+```
+
+**Step 3: Fix all imports**
+
+Same pattern as Task 7: `use crate::X` → `use ml_core::X` for anything from ml-core.
+
+Specific cross-model deps to handle:
+- `tft` imports `crate::liquid::FixedPoint` → `crate::liquid::FixedPoint` (both in ml-supervised ✓)
+- `mamba` imports `crate::liquid::FixedPoint` → `crate::liquid::FixedPoint` (both in ml-supervised ✓)
+- `tft` imports `crate::inference::RealInferenceError` → This module stays in ml facade! Need to either:
+  - Move `RealInferenceError` type to ml-core, OR
+  - Remove this import if it's only used in a few places and can be refactored
+- All models import `crate::dqn::mixed_precision::training_dtype` → `ml_core::mixed_precision::training_dtype`
+- All models import `crate::training::unified_trainer::*` → `ml_core::training::unified_trainer::*`
+
+**Step 4: Update ml/src/lib.rs**
+
+```rust
+pub use ml_supervised::tft;
+pub use ml_supervised::mamba;
+pub use ml_supervised::liquid;
+pub use ml_supervised::tgnn;
+pub use ml_supervised::tlob;
+pub use ml_supervised::kan;
+pub use ml_supervised::xlstm;
+pub use ml_supervised::diffusion;
+pub use ml_supervised::transformers;
+pub use ml_supervised::flash_attention;
+```
+
+Add dependency: `ml-supervised = { workspace = true }`
+
+**Step 5: Verify and fix**
+
+Run: `SQLX_OFFLINE=true cargo check -p ml-supervised 2>&1 | tail -10`
+Run: `SQLX_OFFLINE=true cargo check -p ml 2>&1 | tail -10`
+
+**Step 6: Run tests**
+
+Run: `SQLX_OFFLINE=true cargo test -p ml --lib 2>&1 | tail -5`
+
+**Step 7: Commit**
+
+```bash
+git add -A
+git commit -m "refactor(ml): move 8 supervised models to ml-supervised crate
+
+TFT, Mamba, Liquid, TGNN, TLOB, KAN, xLSTM, Diffusion plus shared
+transformers and flash_attention infrastructure."
+```
+
+---
+
+## Phase 5: Build ml-infra (Training Infrastructure)
+
+### Task 10: Write ml-infra Cargo.toml
+
+**Files:**
+- Create: `crates/ml-infra/Cargo.toml`
+
+Dependencies: `ml-core`, `ml-rl`, `ml-supervised` + `argmin` (for hyperopt), model-specific deps.
+
+Feature flags:
+```toml
+[features]
+default = ["cuda"]
+cuda = ["ml-core/cuda", "ml-rl/cuda", "ml-supervised/cuda"]
+```
+
+**Step 1: Commit**
+
+---
+
+### Task 11: Move Infrastructure Modules to ml-infra
+
+**Files:**
+- Move: `crates/ml/src/hyperopt/` → `crates/ml-infra/src/hyperopt/`
+- Move: `crates/ml/src/ensemble/` → `crates/ml-infra/src/ensemble/`
+- Move: `crates/ml/src/trainers/` → `crates/ml-infra/src/trainers/`
+- Move: `crates/ml/src/benchmark/` → `crates/ml-infra/src/benchmark/`
+- Move: `crates/ml/src/deployment/` → `crates/ml-infra/src/deployment/`
+- Move: `crates/ml/src/training_pipeline.rs` → `crates/ml-infra/src/training_pipeline.rs`
+- Move: `crates/ml/src/walk_forward.rs` → `crates/ml-infra/src/walk_forward.rs`
+- Modify: `crates/ml-infra/src/lib.rs`
+- Modify: `crates/ml/src/lib.rs`
+
+**Step 1: Move modules**
+
+```bash
+for mod in hyperopt ensemble trainers benchmark deployment; do
+  mv crates/ml/src/$mod crates/ml-infra/src/
+done
+mv crates/ml/src/training_pipeline.rs crates/ml-infra/src/
+mv crates/ml/src/walk_forward.rs crates/ml-infra/src/
+```
+
+**Step 2: Write ml-infra/src/lib.rs**
+
+```rust
+pub mod hyperopt;
+pub mod ensemble;
+pub mod trainers;
+pub mod benchmark;
+pub mod deployment;
+pub mod training_pipeline;
+pub mod walk_forward;
+```
+
+**Step 3: Fix all imports**
+
+This is the most complex crate — it imports from all three sub-crates:
+- `use crate::dqn::DQN` → `use ml_rl::dqn::DQN`
+- `use crate::dqn::DQNConfig` → `use ml_rl::dqn::DQNConfig`
+- `use crate::ppo::PPO` → `use ml_rl::ppo::PPO`
+- `use crate::tft::*` → `use ml_supervised::tft::*`
+- `use crate::mamba::*` → `use ml_supervised::mamba::*`
+- `use crate::liquid::*` → `use ml_supervised::liquid::*`
+- `use crate::diffusion::*` → `use ml_supervised::diffusion::*`
+- `use crate::kan::*` → `use ml_supervised::kan::*`
+- `use crate::xlstm::*` → `use ml_supervised::xlstm::*`
+- `use crate::tgnn::*` → `use ml_supervised::tgnn::*`
+- `use crate::tlob::*` → `use ml_supervised::tlob::*`
+- `use crate::cuda_pipeline::*` → `use ml_rl::cuda_pipeline::*`
+- All shared items: `use crate::X` → `use ml_core::X`
+
+**Step 4: Update ml/src/lib.rs**
+
+```rust
+pub use ml_infra::hyperopt;
+pub use ml_infra::ensemble;
+pub use ml_infra::trainers;
+pub use ml_infra::benchmark;
+pub use ml_infra::deployment;
+pub use ml_infra::training_pipeline;
+pub use ml_infra::walk_forward;
+```
+
+Add dependency: `ml-infra = { workspace = true }`
+
+**Step 5: Verify and fix**
+
+Run: `SQLX_OFFLINE=true cargo check -p ml-infra 2>&1 | tail -10`
+Run: `SQLX_OFFLINE=true cargo check -p ml 2>&1 | tail -10`
+
+**Step 6: Run tests**
+
+Run: `SQLX_OFFLINE=true cargo test -p ml --lib 2>&1 | tail -5`
+
+**Step 7: Commit**
+
+```bash
+git add -A
+git commit -m "refactor(ml): move hyperopt, ensemble, trainers to ml-infra crate"
+```
+
+---
+
+## Phase 6: Finalize Facade and Verify
+
+### Task 12: Clean Up ml Facade
+
+**Files:**
+- Modify: `crates/ml/src/lib.rs` — final cleanup
+- Modify: `crates/ml/Cargo.toml` — update dependencies
+
+**Step 1: Update ml/Cargo.toml**
+
+- Add: `ml-core`, `ml-rl`, `ml-supervised`, `ml-infra` as dependencies
+- Remove: all dependencies that are now only used by sub-crates
+- Keep: dependencies used by remaining facade modules (`inference.rs`, `validation/`, `risk/`, etc.)
+
+**Step 2: Verify ml/src/lib.rs has complete re-exports**
+
+Every module path that external consumers use (e.g., `ml::dqn::DQNConfig`, `ml::ensemble::EnsembleCoordinator`) must resolve. Create a checklist from the external consumer import scan (Task 1 research).
+
+**Step 3: Fix remaining facade module imports**
+
+Modules still in ml facade (`inference.rs`, `validation/`, `risk/`, `regime/`, etc.) need import updates:
+- `use crate::dqn::` — these now come from `ml_rl` via re-export, so `use crate::dqn::` still works ✓
+- `use crate::tft::` — same via re-export from `ml_supervised` ✓
+- `use crate::ensemble::` — same via re-export from `ml_infra` ✓
+- `use crate::MLError` — from `ml_core` via `pub use ml_core::*` ✓
+
+This should work because the facade re-exports everything. But verify with compilation.
+
+**Step 4: Verify compilation**
+
+Run: `SQLX_OFFLINE=true cargo check -p ml 2>&1 | tail -5`
+
+**Step 5: Commit**
+
+```bash
+git add -A
+git commit -m "refactor(ml): finalize facade with complete re-exports"
+```
+
+---
+
+### Task 13: Full Workspace Verification
+
+**Files:** None modified
+
+**Step 1: Full workspace check**
+
+Run: `SQLX_OFFLINE=true cargo check --workspace 2>&1 | tail -10`
+Expected: 0 errors across all workspace members
+
+**Step 2: Full workspace clippy**
+
+Run: `SQLX_OFFLINE=true cargo clippy --workspace -- -D warnings 2>&1 | tail -20`
+Expected: 0 errors, 0 warnings
+
+**Step 3: Run ml tests**
+
+Run: `SQLX_OFFLINE=true cargo test -p ml --lib 2>&1 | tail -5`
+Expected: Same test count as baseline (2506+), 0 failures
+
+**Step 4: Run sub-crate tests**
+
+Run each in parallel:
+```bash
+SQLX_OFFLINE=true cargo test -p ml-core --lib 2>&1 | tail -5
+SQLX_OFFLINE=true cargo test -p ml-rl --lib 2>&1 | tail -5
+SQLX_OFFLINE=true cargo test -p ml-supervised --lib 2>&1 | tail -5
+SQLX_OFFLINE=true cargo test -p ml-infra --lib 2>&1 | tail -5
+```
+Expected: Tests distributed across sub-crates, total matches baseline
+
+**Step 5: Run downstream service checks**
+
+Run:
+```bash
+SQLX_OFFLINE=true cargo check -p trading_service 2>&1 | tail -3
+SQLX_OFFLINE=true cargo check -p ml_training_service 2>&1 | tail -3
+SQLX_OFFLINE=true cargo check -p backtesting_service 2>&1 | tail -3
+SQLX_OFFLINE=true cargo check -p trading_agent_service 2>&1 | tail -3
+```
+Expected: All compile with 0 errors (they use `ml` facade, re-exports preserve paths)
+
+**Step 6: Verify compile time improvement**
+
+Run: `SQLX_OFFLINE=true cargo build -p ml --timings 2>&1 | tail -5`
+Open the timings HTML report. Compare per-crate times against the baseline 55.1s.
+
+**Step 7: Commit if any fixes were needed**
+
+---
+
+### Task 14: Update common dev-dependency
+
+**Files:**
+- Modify: `crates/common/Cargo.toml`
+
+**Step 1: Update dev-dependency**
+
+`common` has `ml = { path = "../ml" }` as a dev-dependency. Update the path since ml is still at the same location:
+```toml
+ml = { path = "../ml" }  # Should still work — verify
+```
+
+If paths changed, update accordingly.
+
+**Step 2: Verify**
+
+Run: `SQLX_OFFLINE=true cargo test -p common --lib 2>&1 | tail -5`
+
+**Step 3: Commit if changed**
+
+---
+
+### Task 15: Update Memory and Documentation
+
+**Files:**
+- Modify: `/home/jgrusewski/.claude/projects/-home-jgrusewski-Work-foxhunt/memory/MEMORY.md`
+
+**Step 1: Update MEMORY.md**
+
+Update the Architecture section to reflect the new 5-crate ML structure. Update the ML Model Training section with new crate paths. Remove any stale references to `crates/ml/src/dqn/mixed_precision.rs` etc.
+
+**Step 2: Final commit**
+
+```bash
+git add -A
+git commit -m "refactor(ml): complete crate split — 5 crates, parallel compilation
+
+Split the monolithic ml crate (260K lines, 55s compile) into:
+- ml-core: shared types, traits, features, data, checkpoint, safety
+- ml-rl: DQN + PPO + cuda_pipeline
+- ml-supervised: TFT, Mamba, Liquid, TGNN, TLOB, KAN, xLSTM, Diffusion
+- ml-infra: hyperopt, ensemble, trainers, benchmark, deployment
+- ml: thin facade re-exporting everything
+
+Zero API changes for downstream consumers.
+Incremental rebuild after model edit: ~55s → ~10-15s."
+```
+
+---
+
+## Execution Checklist
+
+| Task | Description | Est. Size | Depends On |
+|------|-------------|-----------|------------|
+| 1 | Verify clean baseline | verify | — |
+| 2 | Create sub-crate dirs + workspace | scaffold | 1 |
+| 3 | Extract shared items from dqn/ | refactor | 2 |
+| 4 | Write ml-core Cargo.toml | config | 2 |
+| 5 | Move core modules to ml-core | **large** | 3, 4 |
+| 6 | Write ml-rl Cargo.toml | config | 4 |
+| 7 | Move RL modules to ml-rl | **large** | 5, 6 |
+| 8 | Write ml-supervised Cargo.toml | config | 4 |
+| 9 | Move supervised models | **large** | 5, 8 |
+| 10 | Write ml-infra Cargo.toml | config | 6, 8 |
+| 11 | Move infra modules | **large** | 7, 9, 10 |
+| 12 | Clean up ml facade | medium | 11 |
+| 13 | Full verification | verify | 12 |
+| 14 | Update common dev-dep | small | 12 |
+| 15 | Update memory + final commit | small | 13 |
+
+## Critical Invariants (Verify After Every Task)
+
+1. **No stubs**: Every function must be wired to its real implementation
+2. **No lost code**: Every module in the inventory must be accounted for
+3. **No API changes**: `use ml::X::Y` must work for all external consumers
+4. **Tests pass**: Same test count, 0 failures
+5. **Clippy clean**: 0 errors, 0 warnings
+6. **Feature flags work**: `--features cuda`, `--no-default-features`, `--features financial` all compile