From 3e5c07249a6a97a78f3bc5cf2df567e6587509a0 Mon Sep 17 00:00:00 2001
From: digital-garbage <133010426+digital-garbage@users.noreply.github.com>
Date: Thu, 3 Sep 2026 18:59:47 +0300
Subject: [PATCH] =?UTF-8?q?fix:=20close=20a=20gap=20in=20the=20H3=20disabl?=
 =?UTF-8?q?e=20pass=20=E2=80=94=20output=5Fguidance=20snapshot=20and=20Dyn?=
 =?UTF-8?q?aShift=20capture?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Both were captured "regardless of the toggle" by original design (so
turning guidance on later works immediately), which meant they kept
running on H3 even after output_guidance/dynashift were forced off —
exactly the kind of function the "disable everything unrelated to
REINS" pass was supposed to catch, missed because the capture point
predates any toggle check.

Confirmed live: the user's own log showed "DynaShift negative bank
updated — 1 entry" after the rest of the disable pass had already
landed.
---
 conditioning.py | 4 +++-
 samplers.py     | 7 ++++++-
 2 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/conditioning.py b/conditioning.py
index 9da9ccd5..de01de64 100644
--- a/conditioning.py
+++ b/conditioning.py
@@ -10208,7 +10208,9 @@ class FunPackVideoRefinerV2(FunPackVideoRefiner):
             # discard otherwise. Separate from the value-function block above because
             # wrong_appearance is deliberately NOT reward-admissible there (repair
             # semantics), yet it is the canonical intrusion signal here.
-            if has_previous_run and refinement_key and not learning_profile.get("skip_learning"):
+            # NOT on H3 -- dynashift (the only reader of this bank) is forced off there, and
+            # the sampler no longer captures a pending latent for it either.
+            if has_previous_run and refinement_key and not learning_profile.get("skip_learning") and not _is_h3:
                 try:
                     try:
                         from .negative_memory import consume_pending, is_negative_profile
diff --git a/samplers.py b/samplers.py
index 782e828f..155cdbee 100644
--- a/samplers.py
+++ b/samplers.py
@@ -8266,7 +8266,12 @@ class FunPackLTXAVSceneChainSampler:
             if self._is_nested(_snap):
                 _parts = [t for t in _snap.unbind() if isinstance(t, torch.Tensor) and t.numel() > 0]
                 _snap = max(_parts, key=lambda t: t.numel()) if _parts else None
-            if isinstance(_snap, torch.Tensor):
+            if isinstance(_snap, torch.Tensor) and not self._is_h3:
+                # NOT on H3: this snapshot only ever feeds output_guidance's value function
+                # and DynaShift's negative bank, both forced off for H3 (see self._is_h3
+                # override above) — capturing it here anyway would just be banking video
+                # latents nothing ever reads, the same waste the "disable everything unrelated
+                # to REINS" pass was for.
                 self._save_output_value_snapshot(refinement_key_input, _snap, None)
                 # DynaShift pending candidate: same run/rating pairing as the snapshot, but
                 # the RAW video latent (fp16) — the rating decides whether it becomes a