docs: Phase A verdict — RA-G1 PASS, RA-G3 starvation-limited on clean data; proceed to D2

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/superpowers/specs/2026-06-05-reward-pnl-alignment-restoration.md

@@ -156,3 +156,18 @@ Date: 2026-06-05
 Falsification trigger (viability): RA-G4 < 0.70 → STOP before 1B-C build.
 Going-in-circles guard: this is the alignment branch of foundation-reshape §5.4; if D1+D2 pass RA-G4/G5 but RA-G6 fails at cluster, the encoder/generalization branch is legitimately isolated — NOT another reward reshape.
 ```
+
+---
+
+## §9. Phase A verdict (2026-06-05, executed on clean data post-L0-fix)
+
+Executed on top of the MBP-10 L0 decode fix (`6df3284d0`) with freshly-regenerated
+clean sidecars, mid-smoke seed 42 b=128, `FOXHUNT_PIN_SURFER_SCAFFOLD=1`:
+
+- **RA-G1 PASS** — `surfer_scaffold_weight` = 0.0 across all 2000 rows; `surfer_scaffold_force_pin` = 1.0. The D1 force-pin mechanism is correctly wired.
+- **RA-G2** — DEFERRED to Phase C (the D2 trainer change needs its own determinism check; pin code committed at `d8447475c`).
+- **RA-G3 INCONCLUSIVE (starvation-limited, NOT a clean fail)** — Pearson 0.016, but only because the agent **flat-collapsed**: TRAIN flat fraction 0.80→1.00, EVAL 100% flat, **0 eval trades, eval pnl $0**. No trades → no pnl variance → degenerate correlation. This is the documented b=128 pure-pnl starvation (`pearl_pure_pnl_mode_starves_b16_controllers`; spec §7 risk 5). Cannot be validated on the RTX 3050 (no b=1024).
+
+**Interpretation:** D1-alone produces flat-collapse, exactly as the spec predicted (Phase 1A regressed −15.5%). Notably, pure-pnl flat ($0) beats the misaligned-shaping run (−$10.7M) — removing the anti-aligned hold-bonus stopped the active bleeding, but with no credit assignment (and/or no real edge) the agent simply stops trading. RA-G3/RA-G4 are not decidable locally.
+
+**Decision (user-directed 2026-06-05):** proceed to build **D2 (Phase C, GAE multi-epoch PPO over the existing rollout buffer)** rather than gate on the locally-unmeasurable RA-G4. RA-G4 viability is consciously deferred to the eventual b=1024 cluster run (Phase D). Phase C begins with the C1 rollout-scaffold audit (no speculative code).