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>
be8f3310d5