jgrusewski be8f3310d5 feat: gradient stability + 9-agent audit bug fixes ...

Per-component gradient clipping:
- 2 new CUDA kernels (dqn_clipped_saxpy, dqn_clip_grad)
- CQL gradient isolated into separate scratch buffer
- Budget allocation: C51=70%, CQL=15%, IQN=10%, Ens=5%
- Dynamic budget — inactive components' share goes to C51

Spectral norm extended to all 10 weight matrices:
- Was trunk-only (W_s1, W_s2), now covers all heads
- 16 u/v power iteration buffers, batched GOFF sync-back
- Uniform sigma_max, end-to-end Lipschitz bounded

Bug fixes from 9-agent audit:
- Backtest episode reset: all 8 fields (was 5), max_equity updated before floor check
- gradient_clip_norm unified: 10.0 everywhere (was 10.0 vs 1.0)
- entropy_coefficient: Option<f64> → f64, single default 0.001
- Close price: unwrap_or(1.0) → direct indexing (no silent wrong rewards)
- step_count: only increments during training (was incrementing on inference)
- Quantile loss: single CUDA context (was 3×), silent fallbacks removed
- MaybeNoisyLinear: single-variant enum removed → direct NoisyLinear
- Budget fractions: exported as pub(crate) constants, referenced not hardcoded

Monitoring:
- Per-component gradient Prometheus gauges (C51 raw, combined)
- Diagnostic logging every 1000 steps

Tests: 7 new gradient budget tests, 1 gradient bounds smoke test
All 488 tests pass (359 ml-dqn + 129 ml)

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

2026-03-27 19:52:44 +01:00

..

2026-03-10-zero-cpu-hotpath.md

docs: add zero-CPU DQN training hot path implementation plan

2026-03-10 15:49:40 +01:00

2026-03-11-dqn-factored-actions.md

feat(dqn): enable Branching DQN with 45 factored actions (5×3×3)

2026-03-11 22:00:13 +01:00

2026-03-13-gpu-test-workflow.md

fix(plan): reviewer round 3 — tggn/tgnn module mapping, exit code, canonical imports

2026-03-13 11:51:58 +01:00

2026-03-14-kanidm-netbird-sso.md

docs: add Kanidm SSO + NetBird mesh design spec and implementation plan

2026-03-14 19:41:40 +01:00

2026-03-15-fused-cuda-training.md

refactor(ml): delete mixed_precision module — BF16 unconditional on CUDA

2026-03-16 16:11:48 +01:00

2026-03-16-gpu-iql-her-porting.md

refactor(cuda): eliminate all CPU fallbacks — CUDA mandatory across ML stack

2026-03-16 21:01:28 +01:00

2026-03-17-candle-hard-refactor.md

refactor(cuda): eliminate candle from ml-core, ml-ppo, and 4 thin crates

2026-03-17 22:27:56 +01:00

2026-03-17-cuda-candle-elimination.md

feat(cuda): complete Candle elimination + cudarc 0.19.3 upgrade

2026-03-17 15:13:04 +01:00

2026-03-18-full-gpu-forward-passes.md

audit: annotate all remaining to_vec/memcpy_dtoh — categorized 158 sites

2026-03-18 19:36:41 +01:00

2026-03-18-gpu-dedup-consolidation.md

fix(ml-supervised): resolve all 104 compile errors — clean build

2026-03-18 07:26:56 +01:00

2026-03-18-gpu-resident-pipeline.md

feat: WORKSPACE COMPILES — zero candle, zero errors, GPU-native

2026-03-18 13:31:35 +01:00

2026-03-21-h100-epoch-optimization-phase1.md

docs: Phase 1 implementation plan — eliminate PER CPU roundtrips

2026-03-21 10:26:11 +01:00

2026-03-21-h100-epoch-optimization-phase2.md

docs: Phase 2 plan — cuBLAS batched GEMM + kernel fusion (10.7ms → <1ms/step)

2026-03-21 11:23:09 +01:00

2026-03-21-h100-epoch-optimization-phase3.md

perf: cuBLAS SGEMM pipeline + dead code elimination — 37s → 86ms/epoch (430x)

2026-03-21 15:33:00 +01:00

2026-03-21-h100-epoch-optimization-phase4.md

perf: keep GPU training data resident across epochs — init 73ms → 0ms

2026-03-21 16:29:32 +01:00

2026-03-21-training-profiles.md

docs: training profile configuration system — spec + plan

2026-03-21 23:22:58 +01:00

2026-03-22-gpu-composite-reward.md

plan: GPU composite reward implementation — 7 tasks

2026-03-22 20:35:57 +01:00

2026-03-22-gpu-segment-tree.md

docs: GPU segment tree spec + plan for O(log n) PER sampling

2026-03-22 00:39:34 +01:00

2026-03-23-9-exposure-dual-distributional.md

plan: 9-exposure action space + dual C51/IQN distributional architecture

2026-03-23 09:03:19 +01:00

2026-03-23-stable-distributional-loss.md

plan: three-layer stable distributional loss architecture

2026-03-23 20:42:08 +01:00

2026-03-24-c51-loss-convergence.md

feat: comprehensive DQN training pipeline overhaul — 16 bug fixes, MSE warmup, financial metrics

2026-03-24 21:48:38 +01:00

2026-03-24-financial-metrics-pipeline.md

fix: segment-based trade lifecycle + reversal P&L computation

2026-03-24 23:57:04 +01:00

2026-03-24-master-trading-agent.md

docs: master plan — maximize all 25 DQN features for profitable ES trading

2026-03-24 22:00:45 +01:00

2026-03-24-remaining-gpu-features.md

fix: segment-based trade lifecycle + reversal P&L computation

2026-03-24 23:57:04 +01:00

2026-03-25-boundary-trade-counting.md

fix: backtest evaluator — correct annualization, window sizing, objective calibration

2026-03-25 19:43:26 +01:00

2026-03-25-config-unification.md

docs: config unification plan — single source of truth, 8 tasks

2026-03-25 09:01:12 +01:00

2026-03-25-eliminate-nvrtc.md

docs: plan to eliminate NVRTC — precompile all 25 CUDA kernels at build time

2026-03-25 23:58:30 +01:00

2026-03-25-reward-v6-fixes.md

docs: reward v6 fixes plan

2026-03-25 01:49:53 +01:00

2026-03-25-trade-lifecycle-fix.md

docs: trade lifecycle fix plan — address review feedback

2026-03-25 23:02:39 +01:00

2026-03-26-eval-cleanup.md

docs: eval cleanup spec + plan — remove dead 5-action path, fix train/eval mismatch

2026-03-26 00:43:58 +01:00

2026-03-26-final-fixes-before-h100.md

docs: final fixes plan — 6 issues blocking H100 deployment

2026-03-26 01:06:28 +01:00

2026-03-27-bug-fixes.md

feat: gradient stability + 9-agent audit bug fixes

2026-03-27 19:52:44 +01:00

2026-03-27-gpu-hot-path-performance.md

feat: gradient stability + 9-agent audit bug fixes

2026-03-27 19:52:44 +01:00

2026-03-27-gradient-stability.md

feat: gradient stability + 9-agent audit bug fixes

2026-03-27 19:52:44 +01:00

2026-03-27-hive-findings-fixes.md

fix: hive findings — capital floor, dueling centering, metrics, CVaR objective

2026-03-27 07:56:50 +01:00

2026-03-27-trade-physics-refactor.md

fix: shared trade physics, IQN arg swap, trailing stop, win_rate

2026-03-27 00:47:34 +01:00

2026-03-27-training-kernel-unification.md

refactor: training kernel fully unified with trade_physics.cuh

2026-03-27 08:16:51 +01:00