Three stabilizers applied to the H=600 DQN smoke after initial run showed
unstable training (early_mvmt=2268× at lr=1e-6, NaN at lr=1e-4):
1. Reward normalization (--reward-scale, default 1000)
Rewards divided by scale BEFORE the Munchausen target. TD error
drops from ~1000 (raw reward magnitude at H=600) into O(1) target /
gradient / weight-update scale. Action selection + rollout-R
reporting use ORIGINAL rewards (so rvr math stays correct against
the Task 7c baseline).
2. Target network (--target-update-every, default 10 episodes)
Separate w_target_dev / b_target_dev buffers. Q_next(s') forward
uses target weights; SGD updates online only. Hard-update copies
online → target every K episodes. Breaks the V_soft(s') chase-its-
own-tail divergence of online-only Munchausen.
3. Gradient clipping (--grad-clip, default 1.0)
New `alpha_clip_inplace_kernel` in alpha_linear_q.cu (element-wise
clamp). Applied to dW and db after grad, before SGD. Safety net.
Diagnostic fix: weight_norm was direction-insensitive — orthogonal
rotations don't change ||W||_F, so early_mvmt read ≈0 even when training.
Switched to weight_distance_from_init = ||W_now − W_init||_F +
||b_now − b_init||_F (captures rotation). q_early = q_init + distance
so kernel's |q_early − q_init| / |q_init| ratio = distance / ||W_init||_F.
With lr bumped back up to 1e-4 (default for the stabilized config),
verified at horizon=100, n_episodes=200:
Q_SPREAD_EMA = 23.64 (≥ 0.05) PASS
ACTION_ENTROPY_EMA = 1.86 (≥ 1.0986) PASS
RETURN_VS_RANDOM_EMA = +0.586 (≥ 0.0) PASS
EARLY_Q_MOVEMENT_EMA = 0.0212 (≥ 0.01) PASS
Overall: PASS (H=6000 scale-up VIABLE)
early_mvmt grew monotonically (0.005 → 0.021) across the 200-episode
run — direction-sensitive diagnostic confirms genuine policy learning.
Audit doc docs/isv-slots.md updated per Invariant 7.
Next: H=600 / 1000-episode run on full data; if PASS holds, Task 13
(H=6000 scale-up) unlocks.
8958637c77