docs: fix 22 review issues in CUDA backtest implementation plan
Chunk 1 fixes:
- Task 1: Replace shared-memory epoch state with global memory + __threadfence()
(shared memory is per-block, multi-block launch would read uninitialized shmem)
- Task 2: Move atomicMin_float/atomicMax_float before kernel definition
- Task 3: Add pub getters for private GPU buffers, fix broken pnl_history logic
(pushing mean_reward N times gives zero std → NaN Sharpe)
- Task 4: Add missing accumulate_q_value/read_q_accumulator methods to GpuTrainingGuard
- Task 6: Fix sort_last_dim tuple destructuring, batch 3 to_scalar into single
8-float readback, move function off GpuTrainingGuard to free function
Chunk 2 fixes:
- Task 8: Add shared-memory parallel reduction for drawdown, win_rate, trade_count
(previously only reduced on thread 0's 1/256th data subset)
- Task 9: Remove dead code, fix task cross-references (12→11), implement
actions_history DtoD copy (was TODO → trade count always 0)
- Task 10: Flesh out BacktestMetrics mapping (6 GPU fields → 16 struct fields),
fix cfg compilation with nested block pattern
Chunk 3 fixes:
- Task 12: Replace "follow same pattern" with actual PPO/supervised code,
enumerate all 8 supervised adapter files, add regression→action mapping
- Task 13: Add complete bitonic sort kernel code for VaR/CVaR extraction,
document intentional spec deviation and deferred Phase 3 capabilities
- Task 14: Specify exact binary path (bin/fxt/src/commands/evaluate_baseline.rs)
- Task 15: Replace placeholder test comments with full synthetic-data validation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
dce318b47f