Final commit of the SP4 Layer C close-out sweep (commits 24accea77, 605a8f526,
ca6315860, 1112abc2a). Adds the comprehensive audit grid to dqn-wire-up-audit.md
covering all 10 host-side EMA/compute sites identified during the sweep:
- 4 migrations completed (C1 redesigned + C2 iqn_loss_ema + C3 utilization_ema +
C4 adaptive_clip full chain).
- 5 documented exceptions: host-aggregated PnL/eval-result inputs
(training_sharpe_ema, max_dd_ema, gamma blend) and multi-step host-normalised
composition (HealthEmaTrackers + LearningHealth) where migration crosses the
brief's "architectural redesign" stop-condition (would require porting whole
compute_epoch_financials or learning_health composition to GPU).
- 1 orphan helper flagged for feedback_wire_everything_up (compute_adaptive_tau
q_div_ema — zero production callers).
- 1 NEW VIOLATION discovered during the audit (calibrate_homeostatic_targets
per-step host EMA loop over 6 mapped-pinned slots) deferred for separate
Layer C task.
- 1 misclassification noted (collector::set_utilization_ema is a setter, not
EMA arithmetic).
Sweep verification post-grep:
git grep -nE "= EMA_BETA \*|= \(1\.0 - EMA_BETA\)|= \(1\.0 - \w*ALPHA\) \*
self\.\w*_ema|= \w*_BETA \* self\.\w*_ema" crates/ml/src/cuda_pipeline/
crates/ml/src/trainers/dqn/
→ 1 match remaining (orphan compute_adaptive_tau q_div_ema, documented).
Going forward: zero-tolerance for new host-side EMA/reduction/mean in
SP4-touched code paths per feedback_no_cpu_compute_strict.md (saved 2026-05-01,
sweep date appended to memory entry).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
6a6b58aec5