docs: handoff 2026-08-03b — afternoon wrap; ms_g6_sp1 = production

policy (d2-d4 via one checkpoint + id override)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SuqdnRgbvmFFduhyw3PUpy

Author: blasphemetheus

docs/planning/HANDOFF_2026-08-03b.md

@@ -0,0 +1,68 @@
+# HANDOFF 2026-08-03b — afternoon session wrap (resume point)
+
+Supersedes HANDOFF_2026-08-03.md's next-moves. Everything through
+`0d948d1` pushed; checkpoints ms_g6_{sp1,sp2}, ms_g7_pressure,
+ms_g8_snippets on B2. Nothing running at wrap time.
+
+## The headline: ms_g6_sp1 is the new production policy
+
+`checkpoints/ms_g6_sp1.bin` (multi-delay {2,3} champion recipe):
+- d2 **434.5 c434** / d3 **413.4 c409** (both 3/3 deterministic)
+- d4 **332.4 c313** with `--delay-id-override 3` (the honest id=4 is an
+  UNTRAINED one-hot and scores 71 — never deploy at untrained ids)
+- One checkpoint covers the whole practical netplay band. Next Direct
+  session should run sp1 at --frame-delay 3 (and use override 3 if the
+  analyzer says the peak drifted to 6).
+
+## Task-list sweep results (19 of 20 closed; #12 + #20-theory remain)
+
+- **Grind-6 spacing A/B**: adjacent-shift interference REFUTED; small
+  pools excel at their low rungs; {2,3,4}'s d2 pin was composition, not
+  structure. No one-variable theory survives the 4-arm matrix (#20).
+- **P5 cycles #1+#2** (fight-state gap): whole-rollout mixing = +43%
+  under pressure but DESTROYS the stand gate (72.9 c1); snippet mixing
+  (103 snippets/23k frames via scripts/snippet_mine.exs + the drill's
+  new --snippet-frames) is SAFE (stand 385.5 c374) but no pressure
+  gain. Methodology validated both directions; the gap wants the HUMAN
+  corpus (--replay-dir! GOTCHA #84) and/or richer labels.
+- **#11 stateful**: drift real (-40% at champion scale); resync hybrid
+  REJECTED (periodic state discontinuity worse than smooth drift);
+  windowed stays deploy default; train-unroll-=-step is the real path.
+  En route: sync runner silently dropped --stateful-step (fixed) — the
+  A/B's activation check caught three identical sham arms.
+- **CycleSim gate PASSED** (morning): hysteresis decode + lag-2 =
+  z chains 151 unbroken / zz metronome, soft=0. Trusted Dolphin-free
+  evaluator for the plain lineage.
+- **SAE cross-arch** (P6 closed): same GT features emerge in GRU and
+  Mamba (own_offstage 0.54/0.71, own_shielding 0.60/0.43); shared
+  knockdown deficit ~0.07 = P4's no-tech-reading at feature level.
+- **Init-forensics ledger CLOSED**: every seed fit the labels
+  (failures were closed-loop); probes saturate (controls caught it).
+- **P4 (yesterday, confirmed)**: nobody reads techs; curve = acceptance
+  test for reaction training.
+
+## New ops rules today
+
+- GOTCHA #83: nix-gc ate the un-rooted appimage-run path (evals died
+  127 at midnight); wrapper now GC-rooted via `nix build -o`.
+- GOTCHA #84: human-session replays die with the temp Dolphin dir —
+  ALWAYS pass --replay-dir (the couch corpus was lost to this).
+- Never deploy/eval at an untrained delay-id; override to nearest.
+- Collection runs must pass --temperature (dedup rule).
+- No shift-jitter in SS-on-queue recipes (grind-3).
+
+## Open threads (in value order)
+
+1. **#12 remote-ping measurement** (Bradley-gated): one Direct session
+   with a friend + analyze_qtrace = the real netplay rung. Fold into
+   the ACAB#182 rematch; regime fears are fully retired (lock/load/DPMS
+   all exonerated; the 08-01 gap was the fixed re-inference bug).
+2. **#20 rung-composition theory**: why {2,3} >> at d2 but {2,3,4} >>
+   at d4. Instruments: {3,4} arm; CycleSim drop-traces (needs
+   queue-embed extension); per-rung boundary maps.
+3. Fight-state gap v3: human corpus + possibly non-multishine pressure
+   labels (a defensive expert, or BC from Bradley's own play).
+4. Train-unroll-=-deploy-step (the O(1) inference endgame).
+5. Technique-id conditioning (the "frame-perfect on command" thread):
+   two techniques, one policy, one trigger channel — all machinery
+   exists.