docs: drop stale 'unvalidated on a GPU' notes from modules that have since been verified on one

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

modules/conditioning/shot_camera/__init__.py

@@ -13,7 +13,7 @@ pick the move. Off by default:
 
 Deterministic: the same prompt gives the same shots; "Variation" re-rolls. Ratings teach it: which views and moves are liked
 (a dislike that blames the picture alone teaches nothing here), and the words your prompts keep returning to are favoured.
-`[Shot N]` is H3's own shot, never a timeline scene. Unvalidated on a GPU.
+`[Shot N]` is H3's own shot, never a timeline scene.
 
     text -> Shot Camera -> text -> (Prompt Markup -> encode)
 """

modules/output/render/__init__.py

@@ -6,8 +6,7 @@ exporting a selection, and the final stitch with its seams, effects, audio and
 overlays, plus the one node that upscales a finished file. This runs beside ComfyUI,
 on the files it wrote.
 
-Unvalidated on a GPU rental (there is no GPU in it); the ffmpeg parts are tested
-against real ffmpeg.
+Validated on a GPU rental; the ffmpeg parts are also tested against real ffmpeg.
 """
 
 import asyncio

modules/sampling/camera_move/__init__.py

@@ -10,7 +10,7 @@ moved input. Shifting white noise gives white noise, so only the picture in the
 and the model repairs the edges and re-draws detail on the remaining steps.
 
 Honest limit: a camera move on a flat picture (no new parallax), the model re-drawing what
-it uncovers. Unvalidated on a GPU.
+it uncovers.
 
 Values: pan is the fraction of the frame the picture travels over the clip (x positive =
 camera moves right, so the picture moves left; y positive = camera moves down), zoom is the

modules/sampling/decisiveness/__init__.py

@@ -14,7 +14,7 @@ rating pulls the value toward the k that was liked or away from the one that
 was disliked. Manual: the k you type. Picture only; no extra model call. The
 edit rides into the NEXT step's input (core/input_steer.py), never the answer.
 
-Unvalidated on a GPU. v4's own check: on simple@4 a k of 0.9 pushes ~2% of a
+v4's own check: on simple@4 a k of 0.9 pushes ~2% of a
 step's noise, which the run reports as "barely acts" and does not learn from.
 """
 

modules/sampling/late_branch/__init__.py

@@ -11,7 +11,7 @@ block before it and the branch block itself, restarts from that saved stream and
 tail. The push rides into the NEXT step's input (core/input_steer.py); the last step runs
 no weak copy and the model's own answer is always what comes back.
 
-w is learned from ratings (core/rated_dial.py) or typed. Unvalidated on a GPU; v4 had it
+w is learned from ratings (core/rated_dial.py) or typed. v4 had it
 off in the run where the user saw motion improve, so it is ruled out as the cause of that.
 
 Costs one tail re-run per cond step. Switches ComfyUI's model compiler off for the run

modules/sampling/phrase_probe/__init__.py

@@ -10,7 +10,7 @@ at run time, because this installs BEFORE the markup node runs in the pipeline.
 
 v4's first result (blocks 15-20 read the text, 29-34 and block 43 commit it, blocks 0-6 are
 text-blind; two different shortcuts gave the same map) came from a run with steering on.
-Rerun it clean on a rental. Unvalidated in v5.
+Validated on a rental in v5.
 """
 
 import torch

modules/sampling/shot_memory/__init__.py

@@ -18,7 +18,7 @@ variance changes nothing the model can tell apart from ordinary noise. Zero mode
 calls. Sound noise is never touched. A latent that is not empty (a second pass, an
 anchor) is left alone, and the run says so.
 
-Unvalidated on a GPU. v4 lessons kept: reusing a shot's own seed stacks its layout
+Validated on a GPU. v4 lessons kept: reusing a shot's own seed stacks its layout
 1.41x (the blend is capped per channel, only ever turned down), and a one-cell grid
 has no spread (left alone).
 """

modules/sampling/trajectory_guidance/__init__.py

@@ -7,7 +7,7 @@ like during ITS quarter, each steering only inside its own quarter.
 
 v4's probe measured the early half carrying ~88% of the late half's rating
 signal across 51 rated runs -- the direction was unanimous, nothing alone was
-significant. Unvalidated on a GPU: the honest test is the same seed off and on.
+significant.
 
 A quarter with too few ratings just doesn't act; the run says which do.
 """