docs(ml-alpha): spec + plan for Phase 1B trainer rollout-buffer + GAE refactor
Companion docs for commit 5cd2f8703 (Phase 1B-A foundation) — got missed
in the foundation commit's staging. The spec drives all 5 sub-phases
(1B-A through 1B-E), the plan expands them into atomic tasks.
* docs/superpowers/specs/2026-06-02-trainer-rollout-buffer-gae.md
— Why (math case + Phase 1A regression evidence)
— Architecture decisions (rollout buffer shape, T_rollout, K_ppo, GAE
kernel, multi-epoch PPO, PER reconciliation, CUDA Graph compat,
determinism preservation)
— 6 falsification gates (G1 alignment / G2 V regression Pearson /
G3 sample efficiency / G4 determinism / G5 cross-source / G6 eval pnl)
— Risks + decision tree from Phase 1A verdict
* docs/superpowers/plans/2026-06-02-trainer-rollout-buffer-gae-implementation.md
— Phase 1B-A tasks (now COMPLETE in 5cd2f8703)
— Phase 1B-B through 1B-E task outlines (expand when prerequisites pass)
Per feedback_investigation_first_falsification_methodology: spec-then-
plan-then-implement; predetermined falsification gates at every phase.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
969caf26c3