feat(dqn-v2): B.3 + C.5 GPU-only replay seed warm-start + CQL α ramp
Plan 3 Tasks 8 + 9. Single commit because Task 9 directly consumes Task 8's
seed-fraction signal; no useful intermediate state.
ISV tail-append:
- [82] SEED_STEPS_TARGET_INDEX — config replay_seed_steps (CPU constructor write)
- [83] SEED_STEPS_DONE_INDEX — GPU-incremented per collect_experiences_gpu
- [84] SEED_FRAC_EMA_INDEX — adaptive EMA of (1 - done/target)
- Fingerprint shifted [80,81] → [85,86]; ISV_TOTAL_DIM 82 → 87
GPU-only design (per user direction "fully gpu driven no cpu involvement"):
- 4 scripted policies as ONE CUDA kernel (scripted_policy_kernel.cu)
- Per-sample policy mix (40% uniform LCG / 20% momentum / 20% mean-rev /
20% vwap-deviation) deterministic by `i % 5`
- Action source switched at launch boundary (CPU per-epoch read of ISV slot
decides which kernel to dispatch; the action computation itself is 100% GPU)
- No CPU physics mirror — existing GPU `experience_env_step` runs unchanged
seed_step_counter_update_kernel.cu:
- Single-thread cold-path; increments DONE, computes FRAC = max(0, 1-done/target)
- Adaptive α matches Task 3/4 convention (α_base × (1 + 0.5×|clamp(sharpe,±2)|))
cql_alpha_seed_update_kernel.cu (Task 9):
- target = config.cql_alpha × max(0, 1 - seed_frac)
- During seed phase (frac=1) → target=0 → CQL α decays to 0 (no pessimism on
exploration data); as frac → 0 → CQL α ramps to config value
- Updates ISV[CQL_ALPHA_INDEX=48]; CQL gradient kernel reads slot 48 via
pinned device-mapped ISV (Plan 1 Task 12 consumer pattern unchanged)
Registry: SEED_STEPS_DONE + SEED_FRAC_EMA both FoldReset; CQL_ALPHA flipped
SchemaContract → FoldReset; SEED_STEPS_TARGET stays SchemaContract.
Read-only monitors (mirror PlanThresholdMonitor / StateKlMonitor pattern):
- monitors/seed_monitor.rs — surfaces ISV[82..85) for HEALTH_DIAG +
controller_activity smoke fire-rate
- monitors/cql_alpha_monitor.rs — surfaces ISV[48] + ISV[84] dependency
Smoke (RTX 3050 Ti, 3 folds × 5 epochs, dqn-smoketest profile with
replay_seed_steps=1000 override so seed phase completes mid-fold):
- All 3 folds saved best-checkpoint
- Fold 2 best Sharpe = 92.4938 at epoch 1 (target range 80-120) ✓
- Per-fold val_metric: f0=3.80 / f1=9.73 / f2=20.24 (loss-based)
- HEALTH_DIAG[3..4] cql_alpha=0.0500 with health=0.49 → base ≈ 0.10 from ISV[48],
consistent with kernel ramping toward final×(1-frac); regime gate
(1-regime)×health applies on top
- 11 cargo check warnings (matches pre-task baseline; no new warnings)
- 6/6 monitor unit tests pass (read/diagnose/observe×fire_rate)
Smoke override rationale: smoke runs ~200 samples per collect (4 episodes ×
50 timesteps) × 5 epochs × 3 folds ≈ 3000 total. Default 100k target would
keep entire smoke in seed phase. Override to 1000 lets the seed→network
transition complete mid-fold so the CQL α ramp is observable.
Per pearl_one_unbounded_signal_per_reward.md: cql_alpha is bounded (clamped
to config_final × (1 - seed_frac) ∈ [0, config_final]), composes safely with
downstream CQL loss.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3cb083f182