diff --git a/docs/plans/2026-02-23-production-hardening-phase2-design.md b/docs/plans/2026-02-23-production-hardening-phase2-design.md
new file mode 100644
index 000000000..f78e4ceb4
--- /dev/null
+++ b/docs/plans/2026-02-23-production-hardening-phase2-design.md
@@ -0,0 +1,199 @@
+# Production Hardening Phase 2 — $100K Live Trading Readiness
+
+> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
+
+**Goal:** Eliminate all remaining safety gaps, wire real data into placeholder paths, verify correctness with integration tests, and establish training infrastructure — making the system ready for $100K live trading.
+
+**Architecture:** Risk-prioritized 4-layer approach: Safety Net (crash prevention) → Correctness (accurate calculations) → Verification (test coverage) → Training Infrastructure (GPU strategy + data pipeline). The liquid-cfc-v2 ensemble extension is integrated as part of Layer 2 correctness work.
+
+**Tech Stack:** Rust (37+ crate workspace), Candle v0.9.1 (ML), Databento (market data), Scaleway (cloud GPU), safetensors (model checkpoints)
+
+---
+
+## Audit Context
+
+### What Was Done (Phase 1)
+- 53-task production hardening merged to main (+2,952/-1,023 lines across 52 files)
+- 80+ TODO/FIXME items resolved, all clippy deny rules enforced
+- 4,234+ tests passing across 30 crate targets, 0 warnings
+
+### What the Audit Found
+Four parallel agents audited the codebase and identified:
+- **92 remaining TODOs** — 4 CRITICAL, 3 HIGH, 8 MEDIUM
+- **7 `std::process::exit()` calls** in temporal_guard.rs (panic-equivalent)
+- **28 critical files with 0 tests**, 49 vacuous `assert!(true)` tests, 107 ignored tests
+- **Model loading has no integrity checks**, ensemble has no per-model circuit breakers
+- **GPU OOM = hard crash** (no runtime VRAM monitoring)
+
+---
+
+## Layer 1 — Safety Net (Crash Prevention)
+
+**Rationale:** These issues can crash the system or cause unrecoverable failures during live trading. Fix first.
+
+### 1.1 Replace `std::process::exit()` in temporal_guard.rs
+- **File:** `ml/src/validation/temporal_guard.rs`
+- **Problem:** 7 calls to `std::process::exit(1)` — kills the entire process without cleanup
+- **Fix:** Replace with `Result`-based error propagation using a new `TemporalGuardError` enum
+- **Risk:** HIGH — process exit during trading = lost positions, no graceful shutdown
+
+### 1.2 Wire real position data into trading agent
+- **File:** `services/trading_agent_service/src/service.rs:804-809`
+- **Problem:** Position data hardcoded to `current_weight: 0.0, current_quantity: 0.0`
+- **Fix:** Query position manager for real portfolio weights and quantities
+- **Risk:** CRITICAL — ensemble allocations are meaningless without real position data
+
+### 1.3 Wire real VaR calculation
+- **Files:** `services/trading_service/src/services/risk.rs:315,415`
+- **Problem:** VaR uses placeholder formula `confidence_level * 1_000_000.0`
+- **Fix:** Wire to the real VaR calculator in `risk/src/var/` which already has parametric, historical, and Monte Carlo methods
+- **Risk:** CRITICAL — risk limits are not enforced if VaR is fake
+
+### 1.4 Kill switch Redis monitoring
+- **File:** `risk/src/safety/kill_switch.rs:378`
+- **Problem:** Redis monitoring task comment says it should be spawned but never is
+- **Fix:** Spawn the monitoring task in `start()`, or document that Redis monitoring is deferred and the local kill switch is sufficient for Phase 1
+- **Risk:** HIGH — distributed kill switch won't propagate across services
+
+### 1.5 GPU OOM detection and CPU fallback
+- **Files:** `ml/src/inference/inference.rs`, `ml/src/inference/inference_engine.rs`
+- **Problem:** GPU out-of-memory = hard crash, no runtime VRAM monitoring
+- **Fix:** Add VRAM usage check before GPU inference, fall back to CPU with warning log if VRAM > 80% threshold
+- **Risk:** HIGH — GPU OOM during live trading = system crash
+
+### 1.6 Per-model circuit breakers in ensemble
+- **File:** `ml/src/inference/inference_ensemble.rs`
+- **Problem:** If one model returns garbage (NaN, extreme values), it contaminates the ensemble vote
+- **Fix:** Add per-model circuit breaker that trips on NaN, repeated identical outputs, or extreme value divergence; ensemble continues with remaining healthy models
+- **Risk:** HIGH — one bad model can cause the entire ensemble to generate bad trades
+
+---
+
+## Layer 2 — Correctness (Accurate Calculations)
+
+### 2.1 Portfolio correlation matrix for Markowitz allocation
+- **File:** `services/trading_agent_service/src/allocation.rs`
+- **Problem:** Uses diagonal covariance (ignores correlations between assets)
+- **Fix:** Implement rolling correlation matrix from historical returns, use in Markowitz optimization
+- **Risk:** MEDIUM — suboptimal allocation but not dangerous (diagonal is conservative)
+
+### 2.2 Feature extraction NaN/Inf guards
+- **File:** `services/trading_service/src/services/enhanced_ml.rs`
+- **Problem:** Feature extraction can produce NaN/Inf from division by zero (e.g., zero volume, zero range), which propagates through model inference
+- **Fix:** Add NaN/Inf check after feature extraction, before normalization. Replace with 0.0 and log warning.
+- **Risk:** HIGH — NaN in model input = NaN in output = unpredictable trades
+
+### 2.3 Model file integrity validation
+- **Problem:** Model files loaded from disk with no integrity verification — corrupted file = silent bad predictions
+- **Fix:** SHA-256 checksum stored alongside safetensors files, verified on load. Schema validation ensures tensor shapes match expected architecture.
+- **Risk:** MEDIUM — unlikely but catastrophic if it happens
+
+### 2.4 Model versioning and rollback
+- **Problem:** No way to roll back to a previous model version if a new one performs poorly
+- **Fix:** Model registry with version tracking, A/B comparison metrics, and one-command rollback
+- **Risk:** MEDIUM — operational risk during model updates
+
+### 2.5 Liquid CfC v2 ensemble integration
+- **Branch:** `worktree-liquid-cfc-v2` (10 commits, +15,014/-2,666 lines)
+- **Status:** LiquidInferenceAdapter and LiquidTrainableAdapter already implemented
+- **Work needed:**
+  - Merge liquid-cfc-v2 branch to main
+  - Register CfC adapter in EnsembleCoordinator (alongside DQN/PPO/TFT/Mamba2)
+  - Add CfC to hyperopt adapter registry
+  - Verify CfC inference latency is within ensemble budget (<10ms)
+  - Add CfC-specific circuit breaker configuration
+- **Risk:** LOW — additive change, ensemble already handles N models
+
+### 2.6 Hyperopt Phase B unblocking
+- **File:** `ml/src/hyperopt/optimizer.rs`
+- **Problem:** Hyperopt works for individual models but multi-model orchestration (Phase B) is blocked
+- **Fix:** Wire ensemble-level hyperopt that optimizes model weights and per-model hyperparameters jointly
+- **Risk:** MEDIUM — training without hyperopt = suboptimal model configurations
+
+---
+
+## Layer 3 — Verification (Test Coverage)
+
+### 3.1 Execution path integration tests (~53 tests)
+- **Target files (0-test critical files):**
+  - `services/trading_service/src/services/risk.rs` (VaR, risk limits)
+  - `services/trading_service/src/services/enhanced_ml.rs` (feature extraction, inference)
+  - `services/trading_agent_service/src/service.rs` (allocation, order generation)
+  - `services/trading_agent_service/src/allocation.rs` (Markowitz optimization)
+  - `services/trading_service/src/core/execution_engine.rs` (order execution)
+- **Approach:** Property-based tests for numerical code, scenario tests for trading logic
+
+### 3.2 Replace vacuous `assert!(true)` tests
+- **Count:** 49 tests that just `assert!(true)` or test trivial construction
+- **Fix:** Replace each with meaningful assertions testing actual behavior
+- **Priority:** Focus on tests in critical paths first (risk, ML, execution)
+
+### 3.3 Integration test: ML → Order → Fill
+- **Scope:** End-to-end test from feature extraction through model inference, ensemble voting, order generation, and simulated fill
+- **Purpose:** Verify the complete trading pipeline produces valid orders from market data
+
+### 3.4 Integration test: Risk cascade → Kill switch
+- **Scope:** Test that risk limit violations properly cascade through circuit breakers to kill switch activation
+- **Purpose:** Verify safety mechanisms actually trigger under stress conditions
+
+---
+
+## Layer 4 — Training Infrastructure
+
+### 4.1 GPU training strategy
+- **Local (RTX 3050 Ti, 4GB VRAM):**
+  - Dev/debug training, small batch sizes (max 230 for PPO)
+  - Rapid iteration on model architecture changes
+  - Feature extraction and data preprocessing
+- **Cloud (Scaleway GPU instances):**
+  - Production training runs with full datasets
+  - Hyperparameter optimization (parallel trials)
+  - Large batch training for all 5 model types
+- **Deliverables:**
+  - Training launcher script that detects GPU and selects local/cloud path
+  - Scaleway instance provisioning configuration
+  - Model artifact sync between local and cloud storage
+
+### 4.2 Data pipeline — Databento OHLCV + MBP-10
+- **Data source:** Databento API for historical market data
+- **Schemas needed:**
+  - **OHLCV (ohlcv-1m, ohlcv-1h):** Primary training data for all models. ~50MB/symbol/year at 1-min bars
+  - **MBP-10 (market-by-price, 10 levels):** Microstructure features, order book depth. ~2-5GB/symbol/year compressed
+- **Data quantity research:**
+  - Minimum: 2 years OHLCV for regime diversity (bull, bear, sideways, high-vol)
+  - Recommended: 5 years OHLCV + 1 year MBP-10 for microstructure features
+  - Symbols: Start with 5-10 liquid instruments (ES, NQ, CL, GC, EUR/USD equivalent)
+- **Storage:** Databento DBN format, already supported by `dbn_sequence_loader.rs` and `dbn_data_source.rs`
+- **Pipeline:**
+  - Databento API client for historical data download
+  - DBN file management (metadata caching already implemented)
+  - Train/validation/test temporal splits (temporal_guard.rs enforces no leakage)
+
+### 4.3 Training pipeline safety
+- **Checkpointing:** Already implemented (safetensors), verify all 5 model types checkpoint correctly
+- **NaN detection:** Add gradient NaN checks during training, auto-halt with last good checkpoint
+- **LR scheduling:** Implement cosine annealing with warmup for production training runs
+- **Metric logging:** Training metrics to structured logs for monitoring dashboards
+- **Reproducibility:** Seed management for all random operations (data shuffling, model init, exploration)
+
+---
+
+## Scope Exclusions
+
+- **Broker integration:** Separate workstream, not blocked by this work
+- **FIX protocol production config:** Phase 2 of broker_gateway_service
+- **Web dashboard enhancements:** Already functional, not in scope
+- **TLI replacement:** Already completed (deleted in Phase 1)
+
+## Success Criteria
+
+1. Zero `std::process::exit()` calls in the codebase
+2. All placeholder data replaced with real calculations (VaR, positions, features)
+3. GPU OOM handled gracefully with CPU fallback
+4. Ensemble survives individual model failures (circuit breakers)
+5. 5 model types (DQN, PPO, TFT, Mamba2, CfC) all in ensemble
+6. >90% code coverage on critical trading paths
+7. End-to-end integration tests pass (ML→Order→Fill, Risk→KillSwitch)
+8. Training runs complete successfully on both local GPU and Scaleway cloud
+9. Databento data pipeline downloads, processes, and feeds into training
+10. Model checksums verified on every load