fix: IQN 4-branch runtime + order counterfactual + per-branch entropy + histogram
IQN kernel rewrite (critical — was corrupting 75% of gradient budget):
- Eliminate compile-time BRANCH_*_SIZE defines (BRANCH_0_SIZE was 5, should be 3)
- All 9 IQN kernels now take runtime b0/b1/b2/b3 params
- 4-branch forward, backward, decode (was 3-branch, missing magnitude)
- branch_actions buffer B*3 → B*4
Action diversity infrastructure:
- 3-cycle counterfactual: dir mirror / mag relabel / order-type cost delta (50× amplified)
- Per-branch entropy: order (d==2) gets 3× boost in C51 + MSE grad kernels
- Position histogram expanded 6→12 bins (all 4 branches get entropy bonus)
- Boltzmann temperature floor raised to 0.5 for order + urgency branches
- Smoke test epochs 3→10 for diversity verification
Note: experience collector still needs bottleneck forward path — currently reads
trainer's bottleneck-layout weights with state_dim K, producing misaligned Q-values.
Next: eliminate weight copy, use direct pointer views into trainer's params_buf.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
dd6143936e