diff --git a/pmoves/creator/installers/MINIMAX-H3-AUTO_INSTALL-RUNPOD.sh b/pmoves/creator/installers/MINIMAX-H3-AUTO_INSTALL-RUNPOD.sh
new file mode 100644
index 0000000000..c1fbaf5769
--- /dev/null
+++ b/pmoves/creator/installers/MINIMAX-H3-AUTO_INSTALL-RUNPOD.sh
@@ -0,0 +1,115 @@
+#!/usr/bin/env bash
+# MiniMax-H3 (NVFP4) — models + aux nodes installer (RunPod / Linux, cu128).
+# Companion to MINIMAX-H3-MODELS-NODES_INSTALL.bat (Windows/5090).
+#
+# H3 nodes are NATIVE in the PMOVES-Creator fork (comfy_extras/nodes_minimax_h3.py)
+# and load in stock ComfyUI — nothing to clone for H3 itself. Only 3 workflow-
+# support node packs are cloned. Loaders are dropdowns → files keep real repo
+# names (no renaming). NVFP4 UNETs halve the diffusion-model size on Blackwell
+# (10.86 vs 20.94 GiB per UNET); peak VRAM is SEQUENTIAL residency, fits 32 GB.
+# Total ≈ 55.5 GB (63 GB with the optional tail).
+#
+# !! ENCODER IS AN ABLITERATED / UNCENSORED FINE-TUNE — no stock alternative
+# !! exists. For UNFCU / client-facing work this is an operator decision (README).
+#
+# RUNTIME-AGNOSTIC: set COMFY_ROOT to target a specific ComfyUI install, else run
+# from inside the ComfyUI root (needs models/ and custom_nodes/).
+#   Options: H3_PROFILE=NVFP4|NVFP4-HQ   INSTALL_TAIL=0|1
+set -euo pipefail
+
+# ───────────────────── Config (override via env) ─────────────────────
+COMFY_ROOT="${COMFY_ROOT:-$PWD}"
+PYTHON_BIN="${PYTHON_BIN:-python3}"
+VENV_DIR="${VENV_DIR:-venv}"
+WANT_TORCH_STACK="${WANT_TORCH_STACK:-auto}"   # auto | cu128 | keep
+CUDA_TAG="${CUDA_TAG:-cu128}"
+TORCH_VERSION="${TORCH_VERSION:-2.8.0}"
+TORCHVISION_VERSION="${TORCHVISION_VERSION:-0.23.0}"
+TORCHAUDIO_VERSION="${TORCHAUDIO_VERSION:-2.8.0}"
+TORCH_INDEX="https://download.pytorch.org/whl/${CUDA_TAG}"
+H3_PROFILE="${H3_PROFILE:-NVFP4}"              # NVFP4 (8-12GB) | NVFP4-HQ (16-24GB)
+INSTALL_TAIL="${INSTALL_TAIL:-0}"
+
+export PIP_DISABLE_PIP_VERSION_CHECK=1 PIP_ROOT_USER_ACTION=ignore
+export PYTHONUNBUFFERED=1 HF_HUB_ENABLE_HF_TRANSFER=1
+
+# ── VERIFIED SOURCES (hf_fs byte-exact, 2026-08-09 recon) ──
+REPO_QUANTS="DmitryDB/MiniMax-H3-ComfyUI-Quants"
+REPO_ENCODER="OTMFLY/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot"
+ENC_MAIN="qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors"
+ENC_TAIL="qwen3vl_32b_minimax_h3_generation_tail_50_63_int8_convrot.safetensors"
+
+# ───────────────────── Helpers ─────────────────────
+die() { echo "[ERROR] $*"; exit 1; }
+need_pkg() { command -v "$1" &>/dev/null || { apt-get update -y && apt-get install -y "$@"; }; }
+get_node() {  # get_node <folder> <git url>
+  local dir=$1 url=$2
+  if [[ -d "custom_nodes/$dir" ]]; then echo " [SKIP] $dir"; else
+    echo " • cloning $dir"; git clone "$url" "custom_nodes/$dir"; fi
+  [[ -f "custom_nodes/$dir/requirements.txt" ]] && \
+    "$PIP" install --no-input --prefer-binary -r "custom_nodes/$dir/requirements.txt" || true
+}
+hf_get() {  # hf_get <repo> <repo-relative-file> <dest-subdir>
+  local repo=$1 file=$2 sub=$3 dir="$COMFY_ROOT/$sub" base
+  base="$(basename "$file")"
+  [[ -f "$dir/$base" ]] && { echo " • $base present — skip"; return 0; }
+  mkdir -p "$dir"
+  echo " • downloading $file"
+  hf download "$repo" "$file" --local-dir "$dir"
+  # hf preserves the repo path — flatten the leaf up, drop the empty subdir
+  if [[ "$file" != "$base" && -f "$dir/$file" ]]; then
+    mv -f "$dir/$file" "$dir/$base"
+    rmdir "$dir/$(dirname "$file")" 2>/dev/null || true
+  fi
+  return 0
+}
+
+# ───────────────────── Verify root + base tools ─────────────────────
+cd "$COMFY_ROOT"
+[[ -d "models" && -d "custom_nodes" ]] || die "COMFY_ROOT=$COMFY_ROOT is not a ComfyUI root."
+echo "[INFO] ComfyUI root: $COMFY_ROOT | profile=$H3_PROFILE tail=$INSTALL_TAIL"
+need_pkg curl git git-lfs
+git lfs install || true
+
+# ───────────────────── Venv + torch ─────────────────────
+[[ -d "$VENV_DIR" ]] || { echo "Creating venv → $VENV_DIR"; $PYTHON_BIN -m venv "$VENV_DIR"; }
+# shellcheck disable=SC1091
+source "$VENV_DIR/bin/activate"
+PIP="$(command -v pip)"
+"$PIP" install --no-input -U pip setuptools wheel "huggingface_hub[cli]" hf_transfer
+
+GPU="$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -n1 || true)"
+[[ "$WANT_TORCH_STACK" == "auto" ]] && { [[ -n "$GPU" ]] && WANT_TORCH_STACK="cu128" || WANT_TORCH_STACK="keep"; }
+echo "[INFO] GPU=${GPU:-none}  WANT_TORCH_STACK=$WANT_TORCH_STACK"
+if [[ "$WANT_TORCH_STACK" != "keep" ]]; then
+  "$PIP" install --no-input --upgrade-strategy only-if-needed \
+    --index-url "$TORCH_INDEX" --extra-index-url https://pypi.org/simple \
+    "torch==${TORCH_VERSION}+${CUDA_TAG}" \
+    "torchvision==${TORCHVISION_VERSION}+${CUDA_TAG}" \
+    "torchaudio==${TORCHAUDIO_VERSION}+${CUDA_TAG}"
+fi
+
+# ───────────────────── Nodes (3 support packs; H3 is native) ─────────────────────
+echo "Cloning workflow-support custom nodes…"
+get_node "rgthree-comfy"            "https://github.com/rgthree/rgthree-comfy"
+get_node "ComfyUI-KJNodes"          "https://github.com/kijai/ComfyUI-KJNodes"
+get_node "ComfyUI-VideoHelperSuite" "https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite"
+
+# ───────────────────── Models ─────────────────────
+echo "Downloading MiniMax-H3 models (profile=$H3_PROFILE)…"
+hf_get "$REPO_QUANTS"  "FL2VA/MiniMax-H3_FL2VA-${H3_PROFILE}.safetensors"  "models/diffusion_models"
+hf_get "$REPO_QUANTS"  "Ref2VA/MiniMax-H3_Ref2VA-${H3_PROFILE}.safetensors" "models/diffusion_models"
+hf_get "$REPO_QUANTS"  "vae/MiniMax-H3_VideoVAE-FP16.safetensors"          "models/vae"
+hf_get "$REPO_QUANTS"  "vae/MiniMax-H3_AudioVAE-FP32.safetensors"          "models/vae"
+echo " [REQUIRED] conditioning encoder (layers 0-49, ~24.6 GiB)"
+hf_get "$REPO_ENCODER" "$ENC_MAIN" "models/text_encoders/MiniMax-H3"
+if [[ "$INSTALL_TAIL" == "1" ]]; then
+  echo " [OPTIONAL] prompt-enhancement tail (layers 50-63, ~7.1 GiB)"
+  hf_get "$REPO_ENCODER" "$ENC_TAIL" "models/text_encoders/MiniMax-H3"
+else
+  echo " [SKIP] optional tail (set INSTALL_TAIL=1 to fetch)"
+fi
+
+echo
+echo "✅ MiniMax-H3 install complete."
+echo "   Encoder dir: models/text_encoders/MiniMax-H3  (CLIPLoader type: minimax)"
diff --git a/pmoves/creator/installers/MINIMAX-H3-MODELS-NODES_INSTALL.bat b/pmoves/creator/installers/MINIMAX-H3-MODELS-NODES_INSTALL.bat
new file mode 100644
index 0000000000..6d5aab9456
--- /dev/null
+++ b/pmoves/creator/installers/MINIMAX-H3-MODELS-NODES_INSTALL.bat
@@ -0,0 +1,189 @@
+@echo off
+setlocal EnableExtensions EnableDelayedExpansion
+chcp 65001 >nul
+
+rem ============================================================================
+rem  MiniMax-H3 (NVFP4) ▸ download models + clone aux nodes  [5090 / Blackwell]
+rem
+rem  H3 = audio+video generation (t2va / fl2va / ref2va). On Blackwell (RTX 5090)
+rem  the NVFP4 (FP4) UNET builds HALVE the diffusion-model size vs INT8-ConvRot
+rem  (10.86 GiB vs 20.94 GiB per UNET — a verified 48% cut). Peak VRAM is
+rem  SEQUENTIAL residency (encoder ~26 GiB is unloaded before the UNET samples),
+rem  not a sum — it fits 32 GB.  Total download ≈ 55.5 GB (63 GB with the
+rem  optional prompt-enhancement tail).
+rem
+rem  The H3 nodes are NATIVE in the PMOVES-Creator fork
+rem  (comfy_extras/nodes_minimax_h3.py) and load in stock ComfyUI — there is NO
+rem  H3 custom node to clone. Loaders are dropdowns, so filenames are free (the
+rem  files keep their real repo names; no renaming). Only the 3 workflow-support
+rem  node packs below are cloned.
+rem
+rem  !! ENCODER IS AN ABLITERATED / UNCENSORED FINE-TUNE (Qwen3-VL-32B ultra-
+rem  !! uncensored-heretic). No stock-Qwen3-VL H3 encoder exists. For UNFCU /
+rem  !! client-facing work this is an EXPLICIT OPERATOR DECISION — see README.
+rem
+rem  RUNTIME-AGNOSTIC: set COMFY_ROOT to target a specific ComfyUI, else run from
+rem  the ComfyUI root (needs models\ and custom_nodes\).
+rem  Options (env):
+rem    COMFY_ROOT     ComfyUI root (default: current dir)
+rem    H3_PROFILE     NVFP4 (default, 8-12GB class) | NVFP4-HQ (16-24GB class)
+rem    INSTALL_TAIL   1 to also fetch the optional 7.6 GB prompt-enhancement tail
+rem ============================================================================
+
+:: ── RESOLVE COMFY ROOT (parametric) ─────────────────────────────
+if defined COMFY_ROOT (set "COMFY=%COMFY_ROOT%") else (set "COMFY=%CD%")
+if not exist "%COMFY%\models" (
+    echo [ERROR] "%COMFY%" is not a ComfyUI root ^(no models\ dir^). Set COMFY_ROOT or cd there.
+    pause & exit /b 1
+)
+if not exist "%COMFY%\custom_nodes" (echo [ERROR] "%COMFY%" has no custom_nodes\ dir. & pause & exit /b 1)
+echo [INFO] ComfyUI root: %COMFY%
+
+if not defined H3_PROFILE set "H3_PROFILE=NVFP4"
+if not defined INSTALL_TAIL set "INSTALL_TAIL=0"
+echo [INFO] UNET profile: %H3_PROFILE%   optional tail: %INSTALL_TAIL%
+
+:: ── PYTHON (ComfyUI portable) — used to bootstrap the hf CLI ─────
+set "PY=%COMFY%\..\python_embeded\python.exe"
+set "PY_SCRIPTS="
+if exist "%PY%" (
+    rem  ~f / ~dp both resolve the ..\ segment, so Scripts\ below is absolute.
+    for %%P in ("%PY%") do (set "PY=%%~fP" & set "PY_SCRIPTS=%%~dpPScripts")
+) else (
+    set "PY=python"
+)
+echo [INFO] Python: !PY!
+
+git --version >nul 2>&1 || (echo [ERROR] Git not in PATH – install Git for Windows. & pause & exit /b 1)
+
+:: ── hf CLI — resolve an ABSOLUTE path, never a bare `hf` ─────────
+:: pip drops console scripts into the TARGET interpreter's Scripts\ dir, and a
+:: stock ComfyUI portable never puts python_embeded\Scripts on PATH. So a bare
+:: `hf download` fails with "not recognized" in this installer's PRIMARY scenario
+:: — even though the pip install immediately below has just succeeded.
+call :resolve_hf
+if not defined HF (
+    echo [INFO] Installing huggingface_hub CLI + hf_transfer into "!PY!"...
+    "!PY!" -m pip install -U "huggingface_hub[cli]" hf_transfer
+    if errorlevel 1 (echo [ERROR] pip install of huggingface_hub failed. & pause & exit /b 1)
+    call :resolve_hf
+)
+if not defined HF (
+    echo [ERROR] hf CLI not found after install.
+    echo         Looked for "!PY_SCRIPTS!\hf.exe" and for hf on PATH.
+    pause & exit /b 1
+)
+echo [INFO] hf CLI: !HF!
+set "HF_HUB_ENABLE_HF_TRANSFER=1"
+rem  For gated/rate-limited pulls authenticate once first:  "%HF%" auth login
+
+:: ── VERIFIED SOURCES (hf_fs byte-exact, 2026-08-09 recon) ───────
+set "REPO_QUANTS=DmitryDB/MiniMax-H3-ComfyUI-Quants"
+set "REPO_ENCODER=OTMFLY/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot"
+set "ENC_MAIN=qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors"
+set "ENC_TAIL=qwen3vl_32b_minimax_h3_generation_tail_50_63_int8_convrot.safetensors"
+
+:: ── CLONE ONLY THE 3 WORKFLOW-SUPPORT NODE PACKS ────────────────
+echo(
+echo -------- Custom nodes ^(H3 nodes are native — not cloned^) --------
+pushd "%COMFY%\custom_nodes"
+call :get_node "rgthree-comfy"            "https://github.com/rgthree/rgthree-comfy"
+call :get_node "ComfyUI-KJNodes"          "https://github.com/kijai/ComfyUI-KJNodes"
+call :get_node "ComfyUI-VideoHelperSuite" "https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite"
+popd
+
+:: ── DOWNLOAD MODELS via hf (real repo names; flattened into subdir) ─
+echo(
+echo -------- Downloading MiniMax-H3 models --------
+:: Every download is checked. A failure here — auth, disk exhaustion, a bad
+:: profile name, a dropped connection mid-pull — must NOT fall through to the
+:: "models ready" banner below: a false success on a 55-63 GB install is
+:: expensive to discover later, usually at generation time.
+call :hf "%REPO_QUANTS%"  "FL2VA/MiniMax-H3_FL2VA-%H3_PROFILE%.safetensors"  "models\diffusion_models"
+if errorlevel 1 goto :download_failed
+call :hf "%REPO_QUANTS%"  "Ref2VA/MiniMax-H3_Ref2VA-%H3_PROFILE%.safetensors" "models\diffusion_models"
+if errorlevel 1 goto :download_failed
+call :hf "%REPO_QUANTS%"  "vae/MiniMax-H3_VideoVAE-FP16.safetensors"          "models\vae"
+if errorlevel 1 goto :download_failed
+call :hf "%REPO_QUANTS%"  "vae/MiniMax-H3_AudioVAE-FP32.safetensors"          "models\vae"
+if errorlevel 1 goto :download_failed
+echo   [REQUIRED] conditioning encoder (layers 0-49, ~24.6 GiB)
+call :hf "%REPO_ENCODER%" "%ENC_MAIN%" "models\text_encoders\MiniMax-H3"
+if errorlevel 1 goto :download_failed
+if "%INSTALL_TAIL%"=="1" (
+    rem  Parens MUST be escaped inside a block: an unescaped ^) closed this if-body
+    rem  early, so the 7.1 GiB tail downloaded unconditionally and "[SKIP] optional
+    rem  tail" printed right after it. Verified before/after by running the script.
+    echo   [OPTIONAL] prompt-enhancement tail ^(layers 50-63, ~7.1 GiB^)
+    call :hf "%REPO_ENCODER%" "%ENC_TAIL%" "models\text_encoders\MiniMax-H3"
+    if errorlevel 1 goto :download_failed
+) else (
+    echo   [SKIP] optional tail ^(set INSTALL_TAIL=1 to fetch^)
+)
+
+echo(
+echo -------------------------------------------------------------
+echo   MiniMax-H3 models + support nodes ready.
+echo   Encoder dir: models\text_encoders\MiniMax-H3  (CLIPLoader type: minimax)
+echo -------------------------------------------------------------
+pause
+exit /b 0
+
+
+:download_failed
+echo(
+echo -------------------------------------------------------------
+echo   [ERROR] A model download FAILED — the install is INCOMPLETE.
+echo   Do NOT treat the models as ready. Fix the cause and re-run:
+echo     auth    : "!HF!" auth login   ^(gated or rate-limited repos^)
+echo     disk    : needs ~55.5 GB free ^(63 GB with INSTALL_TAIL=1^)
+echo     profile : H3_PROFILE must be NVFP4 or NVFP4-HQ
+echo     network : transient drops are common on multi-GiB pulls
+echo   Files already fetched are kept, so a re-run resumes.
+echo -------------------------------------------------------------
+pause
+exit /b 1
+
+
+:: ==================== SUBROUTINES ============================
+
+:resolve_hf
+rem  Prefer the CLI that ships next to the interpreter we install INTO — a stock
+rem  ComfyUI portable does not put that Scripts\ dir on PATH. Fall back to PATH
+rem  for venv / system installs where a bare `hf` does resolve.
+set "HF="
+if defined PY_SCRIPTS if exist "%PY_SCRIPTS%\hf.exe" set "HF=%PY_SCRIPTS%\hf.exe"
+if not defined HF (where hf >nul 2>&1 && set "HF=hf")
+goto :eof
+
+:get_node
+set "DIR=%~1"
+set "URL=%~2"
+if not exist "%DIR%" (echo   • cloning %DIR% & git clone "%URL%" "%DIR%") else (
+    echo   • updating %DIR%
+    if exist "%DIR%\.git" (pushd "%DIR%" & git pull --ff-only & popd) else (echo     [WARN] %DIR% not a git repo – skip)
+)
+if exist "%DIR%\requirements.txt" ("%PY%" -m pip install --upgrade -r "%DIR%\requirements.txt")
+goto :eof
+
+:hf
+rem  %1 = repo id   %2 = repo-relative file (fwd slashes)   %3 = dest subdir (rel to COMFY)
+set "H_REPO=%~1"
+set "H_FILE=%~2"
+set "H_DIR=%COMFY%\%~3"
+set "H_WIN=%H_FILE:/=\%"
+for %%A in ("%H_WIN%") do set "H_BASE=%%~nxA"
+if exist "%H_DIR%\%H_BASE%" (echo   • %H_BASE% already present – skip & exit /b 0)
+if not exist "%H_DIR%" mkdir "%H_DIR%"
+echo   • downloading %H_FILE%
+"%HF%" download "%H_REPO%" "%H_FILE%" --local-dir "%H_DIR%"
+rem  Propagate, do not warn-and-continue: the caller aborts the whole install.
+if errorlevel 1 (echo     [ERROR] Download failed: %H_FILE% & exit /b 1)
+rem  hf preserves the repo path — flatten the leaf up into H_DIR, drop empty subdir
+if not "%H_WIN%"=="%H_BASE%" (
+    if exist "%H_DIR%\%H_WIN%" (
+        move /Y "%H_DIR%\%H_WIN%" "%H_DIR%\%H_BASE%" >nul
+        for %%D in ("%H_WIN%") do rd "%H_DIR%\%%~pD" 2>nul
+    )
+)
+goto :eof
diff --git a/pmoves/creator/installers/MINIMAX-H3-README.md b/pmoves/creator/installers/MINIMAX-H3-README.md
new file mode 100644
index 0000000000..97efb0a73d
--- /dev/null
+++ b/pmoves/creator/installers/MINIMAX-H3-README.md
@@ -0,0 +1,181 @@
+# MiniMax-H3 (NVFP4) ComfyUI Installer — 5090 / Blackwell
+
+Installs the **MiniMax-H3** audio+video generation models (NVFP4 quant) plus the
+3 workflow-support node packs, against **either** runtime:
+
+- the in-tree **PMOVES-Creator** fork (native H3 nodes), or
+- the operator's **Pinokio** ComfyUI fork.
+
+Both scripts are **parametric on `COMFY_ROOT`** (set it, or run from the ComfyUI root):
+
+| Script | Host | Options (env) |
+|--------|------|---------------|
+| `MINIMAX-H3-MODELS-NODES_INSTALL.bat` | Windows (5090) | `H3_PROFILE`, `INSTALL_TAIL`, `COMFY_ROOT` |
+| `MINIMAX-H3-AUTO_INSTALL-RUNPOD.sh` | Linux / RunPod | same + torch-stack pins |
+
+> Sources below were **byte-exact verified** against the HF file index on
+> 2026-08-09 (`hf_fs` listings + each repo's `SHA256SUMS`). Where a claim could
+> not be confirmed it is marked UNVERIFIED — nothing is guessed.
+
+## Why NVFP4 (Blackwell)
+
+The RTX 5090 (Blackwell, sm_120) has native FP4. The **NVFP4** UNET builds are
+**10.86 GiB** each vs **20.94 GiB** for INT8-ConvRot — a verified **48% cut**,
+consistent with the halving claim (block-scaled NVFP4, not AWQ; covers all 208
+main + token-refiner matrices). Peak VRAM is **sequential residency** — the
+~24.6 GiB encoder is unloaded before the UNET samples, so it is **not** a sum and
+fits 32 GB.
+
+## H3 nodes are NATIVE — only 3 packs are cloned
+
+The H3 nodes ship in the PMOVES-Creator fork at `comfy_extras/nodes_minimax_h3.py`
+(`EmptyMiniMaxH3LatentAV`, `MiniMaxH3ImageToVideo` = t2va/fl2va,
+`MiniMaxH3ReferenceToVideo` = ref2va, `MiniMaxH3SigmaShift`); INT8/ConvRot layouts
+live in `comfy/quant_ops.py`. They load in **stock ComfyUI** (commit `14b05228`)
+with no core patch. **There is no H3 custom node to install.** The installer clones
+only the workflow-support packs:
+
+- `rgthree-comfy` · `ComfyUI-KJNodes` · `ComfyUI-VideoHelperSuite`
+
+## Filenames are FREE — no renaming
+
+ComfyUI's H3 loaders select from **dropdowns**, so the on-disk filename does not
+matter. The installer keeps each file's **real repo name** and only flattens the
+`hf download` repo-subpath into the target `models/` subdir.
+
+> **Do not** chase the `minimax_h3_fl2va_pruned_int8_convrot.safetensors` name from
+> the original brief — that string is **UNVERIFIED / likely nonexistent**. The
+> `*_pruned_*` naming family traces to a *different, structurally different* repo
+> (`Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI`, pruned **W4A8**). Do not mix
+> families. DmitryDB's NVFP4 files retain all 50 transformer blocks (not pruned).
+
+## Verified model sources
+
+### UNETs + VAEs — `DmitryDB/MiniMax-H3-ComfyUI-Quants`
+
+| File (repo path) | Size | → dir |
+|---|---:|---|
+| `FL2VA/MiniMax-H3_FL2VA-NVFP4.safetensors` | 10.862 GiB | `models/diffusion_models` |
+| `Ref2VA/MiniMax-H3_Ref2VA-NVFP4.safetensors` | 10.862 GiB | `models/diffusion_models` |
+| `vae/MiniMax-H3_VideoVAE-FP16.safetensors` | 4.850 GiB | `models/vae` |
+| `vae/MiniMax-H3_AudioVAE-FP32.safetensors` | 0.564 GiB | `models/vae` |
+
+Profiles present: `NVFP4` (10.86 GiB, 8–12 GB class), `NVFP4-HQ` (13.60 GiB, 16–24 GB
+class), and three `INT8-ConvRot` profiles (20–22 GiB, RTX 30/40). Set `H3_PROFILE`
+to switch (default `NVFP4`). License on these files is the **MiniMax-H3 community
+license** (not Apache).
+
+### Encoder — `OTMFLY/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot`
+
+| File | Size | → dir | Status |
+|---|---:|---|---|
+| `qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors` | 24.553 GiB | `models/text_encoders/MiniMax-H3` | **REQUIRED** |
+| `qwen3vl_32b_minimax_h3_generation_tail_50_63_int8_convrot.safetensors` | 7.086 GiB | `models/text_encoders/MiniMax-H3` | OPTIONAL (`INSTALL_TAIL=1`) |
+
+Selected in **CLIPLoader** with type `minimax`. No NVFP4 build of the encoder
+exists — it is INT8 at 24.55 GiB regardless.
+
+> ### CORRECTION — the layers-50–63 "tail" is NOT the conditioning encoder
+> The brief assumed the ~7.6 GB tail "may be all H3 needs" (saving ~19 GB). **That
+> is wrong.** The repo README states H3 "consumes the unnormalized hidden state
+> after language layer 49" — so the **layers 0–49 file (24.55 GiB) is REQUIRED**.
+> The tail (layers 50–63 + final norm + LM head) is used **only** for optional
+> prompt enhancement (loaded temporarily by the `ComfyUI-MiniMax-H3-Guide` node,
+> then unloaded). Shipping only the tail produces a workflow that cannot encode a
+> prompt. **Budget the 24.55 GiB file.**
+
+## ⚠️ Operator decision — abliterated / uncensored encoder
+
+Both candidate encoder repos derive from `llmfan46/Qwen3-VL-32B-Instruct-ultra-
+uncensored-heretic` (an **abliterated / uncensored** fine-tune). **No stock
+Qwen3-VL H3 encoder exists.** For PMOVES / UNFCU / client-facing work this is an
+**explicit operator decision** — confirm it is acceptable before deploying.
+
+**Alternative source (recommended upstream):** `ethanfel/Qwen3-VL-32B-Ultra-Heretic-
+H3-ComfyUI-INT8-ConvRot` — same content, authored by the `ComfyUI-MiniMax-H3-Guide`
+node author (416 likes vs OTMFLY's 4; OTMFLY is a 2-file mirror). Byte sizes and the
+pinned upstream revision match, but SHA-256 was **not** cross-compared between the
+two (UNVERIFIED equivalence). The installer defaults to OTMFLY because its exact
+filenames are byte-confirmed; switch `REPO_ENCODER` to ethanfel if you prefer
+upstream (confirm ethanfel's exact filename first).
+
+## Exact `hf download` lines
+
+```bash
+export HF_HUB_ENABLE_HF_TRANSFER=1
+
+# --- diffusion_models (NVFP4, Blackwell) ---
+hf download DmitryDB/MiniMax-H3-ComfyUI-Quants FL2VA/MiniMax-H3_FL2VA-NVFP4.safetensors  --local-dir models/diffusion_models
+hf download DmitryDB/MiniMax-H3-ComfyUI-Quants Ref2VA/MiniMax-H3_Ref2VA-NVFP4.safetensors --local-dir models/diffusion_models
+
+# --- vae ---
+hf download DmitryDB/MiniMax-H3-ComfyUI-Quants vae/MiniMax-H3_VideoVAE-FP16.safetensors --local-dir models/vae
+hf download DmitryDB/MiniMax-H3-ComfyUI-Quants vae/MiniMax-H3_AudioVAE-FP32.safetensors --local-dir models/vae
+
+# --- text_encoders (REQUIRED: layers 0-49 conditioning encoder, 24.55 GiB) ---
+hf download OTMFLY/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot \
+  qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors \
+  --local-dir models/text_encoders/MiniMax-H3
+
+# --- text_encoders (OPTIONAL: prompt-enhancement tail, layers 50-63, 7.09 GiB) ---
+hf download OTMFLY/Qwen3-VL-32B-Ultra-Heretic-MiniMax-H3-ComfyUI-INT8-ConvRot \
+  qwen3vl_32b_minimax_h3_generation_tail_50_63_int8_convrot.safetensors \
+  --local-dir models/text_encoders/MiniMax-H3
+```
+
+> **hf-download path note:** `hf download` preserves repo-relative paths, so the
+> first four land under `FL2VA/`, `Ref2VA/`, `vae/` subfolders of `--local-dir`
+> (ComfyUI scans recursively, so they still resolve). The `.bat`/`.sh` installers
+> **flatten** the leaf up into the target dir and drop the empty subfolder; the raw
+> lines above do not. `NVFP4-HQ` swap: substitute `-NVFP4-HQ` in the two UNET names.
+
+## Download totals
+
+| Set | GB |
+|---|---:|
+| 2× NVFP4 UNET | 23.33 |
+| 2× VAE | 5.81 |
+| Encoder (0–49, required) | 26.36 |
+| Tail (50–63, optional) | 7.61 |
+| **Working set (no tail)** | **55.50** |
+| Full set (with tail) | 63.11 |
+
+## Run
+
+```bat
+:: Windows (5090)
+set "COMFY_ROOT=D:\path\to\ComfyUI"
+MINIMAX-H3-MODELS-NODES_INSTALL.bat
+```
+
+```bash
+# Linux / RunPod
+COMFY_ROOT=/workspace/ComfyUI bash MINIMAX-H3-AUTO_INSTALL-RUNPOD.sh
+```
+
+## Verify (T4 pass criteria)
+
+- The 3 node packs are present under `custom_nodes/`.
+- The NVFP4 UNETs + VAEs + the 24.55 GiB encoder are downloaded to the dirs above.
+- Both runtimes load the H3 nodes (native — appear in node search).
+- The workflow's loaders resolve (no red "missing file" nodes).
+- A test generation runs within 32 GB VRAM (watch `nvidia-smi`; sequential-residency
+  peak ≈ 26.4 GiB — encoder and UNET are **not** co-resident).
+
+## UNVERIFIED / caveats (from recon)
+
+1. **NVFP4 never run end-to-end on a 5090** — DmitryDB's RTX-50 ratings are
+   architecture-based; all PASS results in their table are INT8 on a 4090. NVFP4
+   passed numerical/structural validation only. Treat first-run success as unproven.
+2. **Uncensored encoder** — operator decision (above); no stock alternative exists.
+3. **OTMFLY vs ethanfel** equivalence is size-based only (SHA not cross-checked).
+4. Optional-tail path needs `github.com/ethanfel/ComfyUI-MiniMax-H3-Guide`
+   (verified stack: comfy-kitchen 0.2.26, comfy-aimdo 0.4.11, PyTorch 2.8.0+cu128;
+   comfy-kitchen recommends CUDA 13.0+ — 12.8 completes with unoptimized kernels).
+
+## Operator inputs still required to reconcile
+
+1. **SEAP `.bat` contents + folder listing** (operator-local on the 5090).
+2. **The H3 workflow JSON** — to confirm which loaders/encoder path it wires.
+3. **Aitrepreneur's model-mirror link** — cross-check against the verified sources
+   above (which stand on their own; the mirror is a confirmation, not a dependency).
diff --git a/pmoves/docs/operations/B850_BRINGBACK_RUNBOOK.md b/pmoves/docs/operations/B850_BRINGBACK_RUNBOOK.md
new file mode 100644
index 0000000000..3ef1342bb1
--- /dev/null
+++ b/pmoves/docs/operations/B850_BRINGBACK_RUNBOOK.md
@@ -0,0 +1,215 @@
+# B850 "Knuckles" Bring-Back Runbook — Tailscale/SSH Re-Enroll + Security Follow-ups
+
+> **Status:** operator-run. Authored by z890-claude 2026-08-09 to close the "not
+> documented" gap — no prior checklist listed the exact pending enroll commands.
+> **Node:** B850 "Knuckles" — primary dev host + heavyweight ROCm inference, and
+> the **single data-tier home** (Postgres / NATS / the JuiceFS metadata store).
+> Do **not** split-brain the data tier onto another node.
+
+B850 is back online after being offline. This runbook re-enrolls it into the
+Tailscale mesh, verifies fleet visibility, and closes two **HIGH-severity**
+security items on its live `juicefs-mount`.
+
+---
+
+## 0. Resolve the hostname FIRST (read-only)
+
+B850 appears under three different names across the docs — they are **not**
+interchangeable for enrollment:
+
+| Name | Source | Authority |
+|------|--------|-----------|
+| `pmoves-b850-ai-top` | `.claude/context/runner-topology.md` | **authoritative** |
+| `pmoves-9850x3d-r9700` | some runbooks | drift |
+| `pmoves-rdna4` | AI-Top service script | drift |
+
+The enroll `DEVICE=` and the `tailscale up --hostname` **must match the name the
+node is actually registered under**. Confirm the live name before doing anything:
+
+```bash
+# On any node already on the tailnet (read-only — changes nothing).
+# fleet-status, not raw `tailscale status`: it prints the hostname column and
+# redacts IPs, which is the standing convention (BOOTSTRAP.md § Fleet view —
+# "never raw tailscale status for public IPs"). Hostnames are all this step needs.
+make -C pmoves fleet-status | grep -iE 'b850|knuckles|9850|rdna4'
+```
+
+Set `CONFIRMED` to whatever that prints (expected: `pmoves-b850-ai-top`) and use
+it verbatim everywhere below. If the node does not appear at all, it has no live
+registration yet — proceed to step 1 (enroll) which creates it.
+
+> Hostname canonicalization across the three docs is flagged as a **separate
+> docs-fix** (deferred this pass) — do not "fix" it by renaming the live node.
+
+---
+
+## 1. Generate the enrollment token (on z890 or any owner node)
+
+```bash
+# Sources CHIT-managed fleet secrets into the tier env files (never edit env.shared).
+make -C pmoves secrets-funnel
+
+# Needs CHIT_PASSPHRASE in the environment (voice-activated / CHIT vault — do NOT
+# paste it on a shared CLI). Target: pmoves/mk/infra.mk : fleet-enroll.
+export CHIT_PASSPHRASE=...              # from the CHIT vault
+make -C pmoves fleet-enroll ROLE=owner DEVICE="$CONFIRMED"
+```
+
+`ROLE=owner` because B850 is your own management/inference workstation, not a
+third party. (Enrollment tokens for `partner`/`guest` are for others — see the
+`enrollment-is-for-others` note.)
+
+---
+
+## 2. Join the tailnet FROM B850 with the correct tags
+
+Run **on B850**. The tags come from `pmoves/configs/tailscale-acl-policy.json`
+(the `tag:lab` fleet-management block, lines ~120–154) — **not** the z890
+launcher script's tag set:
+
+```bash
+sudo tailscale up \
+  --hostname "$CONFIRMED" \
+  --accept-routes \
+  --accept-dns \
+  --advertise-tags=tag:pmoves,tag:gpu,tag:lab
+```
+
+Why these three tags:
+- `tag:pmoves` — full-mesh membership (port-22 reachability, exit-node egress).
+- `tag:gpu` — B850 serves Ollama/llama-server to the fleet (a destination in the
+  `:11434` / gpu-orchestrator rules).
+- `tag:lab` — fleet-management identity; grants root-SSH to the KVM concentrators
+  via the scoped `tag:lab → tag:vps,tag:exit` rule.
+
+> **Do NOT untag B850 to "fix" reachability.** Measured 2026-08-04, untagging a
+> tagged compute node silently drops it out of `tag:gpu` (breaking fleet Ollama
+> at `:11434`) and out of the exit-node egress src list. Untagging pmoves-4090
+> took it offline from the mesh. The ACL, not a tag removal, is the reachability
+> mechanism here.
+
+---
+
+## 3. Approve the tags (admin console, if prompted)
+
+If Tailscale prompts for tag approval (device-auth or tag-owner approval), approve
+`tag:pmoves`, `tag:gpu`, `tag:lab` for the node in the admin console.
+
+---
+
+## 4. Verify mesh + fleet visibility
+
+```bash
+# From any node on the tailnet. fleet-status is the Known Road: it prints
+# hostname / OS / connection and redacts IPs (see BOOTSTRAP.md § Fleet view).
+make -C pmoves fleet-status | grep -i "$CONFIRMED"   # expect: present, online
+
+make -C pmoves fleet-status                          # expect: relay health green
+```
+
+**Tags are not in either of those.** `tailscale status` does not print per-node
+tags at all — its third column shows the *owner*, which for a tagged node reads
+`tagged-devices`, and `fleet-status` does not surface that column either. Tags
+live only in the JSON, so read them there and print nothing but hostname + tags
+(no addresses, so the no-IPs convention still holds):
+
+```bash
+tailscale status --json \
+  | jq -r --arg h "$CONFIRMED" '.Peer[] | select(.HostName==$h) | "\(.HostName): \(.Tags // ["<untagged>"] | join(", "))"'
+# expect: pmoves-b850-ai-top: tag:gpu, tag:lab, tag:pmoves
+```
+
+**T2 pass criteria:** `make -C pmoves fleet-status` shows `$CONFIRMED` online and
+relay health green; the JSON query above lists `tag:pmoves`, `tag:gpu`, `tag:lab`.
+
+---
+
+## 5. SSH — nothing to install on B850
+
+The fleet does **not** install SSH keys on B850. Access is **Tailscale SSH**,
+authorized by tag in the ACL (`tag:lab` sources reach `tag:vps`/`tag:exit` as
+root; owner identities reach nodes per the owner rule). There is no per-node key
+step here.
+
+> **Tracked gap:** the per-node SSH-key story is still incomplete fleet-wide — see
+> `pmoves/docs/handoffs/doc-coordination-juicefs-ingest-2026-08-06.md` (§ around
+> lines 28–31: `claude-pmoves` SSH is authorized only on 4090 + jetsons, not the
+> 5090). This does not block B850 (owner-role, Tailscale-SSH by tag), but keep it
+> in mind when driving *other* nodes from z890.
+
+---
+
+## 6. Security follow-ups — HIGH severity (operator)
+
+B850's currently-running `juicefs-mount` was started with the **Supabase admin
+password inline in the container command line**, so it is visible in `ps` and
+`docker inspect` to any local user — and has been for **days**. Ref:
+`pmoves/docs/handoffs/juicefs-cross-node-storage-blocker-2026-08-04.md` §
+"Security item" (lines ~79–95).
+
+### 6a. Re-create the mount with `META_PASSWORD` (no inline secret)
+
+The password must be passed via the `META_PASSWORD` env var so the DSN in the
+command line carries no credential. The canonical targets already do this:
+
+```bash
+# Local-DB mount on the data-tier host (B850), password via env, cache bounded:
+export SUPABASE_DB_PASSWORD=...          # from the CHIT secrets pipeline
+make -C pmoves juicefs-mount-local
+```
+
+`juicefs-mount-local` (see `pmoves/mk/egress.mk`) now also injects **per-host
+bounded cache flags** (`scripts/juicefs-cache-bounds.sh`) so B850 does not inherit
+the JuiceFS 100 GiB default nor self-disable caching on a full disk. Confirm the
+new container has **no password** in its command line:
+
+```bash
+docker inspect juicefs-mount --format '{{json .Args}}'    # must NOT contain the password
+docker inspect juicefs-mount --format '{{json .Config.Env}}' | grep -c META_PASSWORD   # expect 1
+```
+
+> **This narrows the exposure — it does not close it. Read before signing off.**
+>
+> `docker run -e META_PASSWORD` (value-less, inherited from the shell) makes Docker
+> persist the **expanded** value into the container's `.Config.Env`. So this step
+> moves the credential out of `.Args` — where it was visible in **both** `ps` and
+> `docker inspect` — into `.Config.Env`, where it is still visible to
+> `docker inspect`. The second verification command above is itself the proof: it
+> greps `.Config.Env` and expects a hit.
+>
+> **Who can still read it:** any local user in the `docker` group (and thus root),
+> via `docker inspect juicefs-mount`. That is a smaller set than "anyone who can run
+> `ps`", which is the point of doing this — but it is not "no cleartext in inspect".
+>
+> **Real remediation:** file-mounted Docker secrets with a `*_FILE` indirection read
+> inside the entrypoint, which keeps the value out of `.Args` *and* `.Config.Env`.
+> The fleet already specifies this — see `#2492 § 8` (materialized per node as a
+> file-mounted secret, never committed, never in `ps`), following the `#1901`
+> precedent. Until that lands here, treat § 6a as *reduction*, not closure, and
+> rotate on the assumption the value is readable by the docker group.
+
+### 6b. Rotate the exposed Supabase admin password
+
+Because the old password sat in process listings and container metadata for at
+least three days, **rotate it** after the mount is re-created without it. Rotate
+in Supabase, then flow the new value through the CHIT secrets pipeline
+(`make -C pmoves secrets-funnel`) — do not hand-edit `env.shared`. Any other
+consumer of `SUPABASE_DB_PASSWORD` / the JuiceFS meta DSN (the gateway, cross-node
+mounts) picks up the new value on next `up`.
+
+**T2 security pass criteria:** `juicefs-mount` re-created with `META_PASSWORD`
+(no cleartext in `ps`, and none in `.Args` — but **still readable in
+`.Config.Env`** via `docker inspect`, see the note in § 6a); Supabase admin
+password rotated and funneled. The file-mounted-secret work in `#2492 § 8` is
+what closes `inspect`; this pass does not.
+
+---
+
+## Related
+
+- Cross-node mounts on 4090/5090 depend on B850 being up + reachable →
+  `JUICEFS_CROSS_NODE_MOUNT_RUNBOOK.md`.
+- Cache-bounds helper (wired into all mount call sites) →
+  `pmoves/scripts/juicefs-cache-bounds.sh`.
+- Storage-backend blocker + security note →
+  `pmoves/docs/handoffs/juicefs-cross-node-storage-blocker-2026-08-04.md`.
diff --git a/pmoves/docs/operations/JUICEFS_CROSS_NODE_MOUNT_RUNBOOK.md b/pmoves/docs/operations/JUICEFS_CROSS_NODE_MOUNT_RUNBOOK.md
new file mode 100644
index 0000000000..280d9dc121
--- /dev/null
+++ b/pmoves/docs/operations/JUICEFS_CROSS_NODE_MOUNT_RUNBOOK.md
@@ -0,0 +1,120 @@
+# JuiceFS Cross-Node Mount Runbook — 4090 / 5090
+
+> **Status:** operator-run. Authored by z890-claude 2026-08-09.
+> **Goal:** mount the shared JuiceFS filesystem on the 4090 and 5090 so both can
+> read/write fleet content, with a **per-host bounded cache** so neither inherits
+> the JuiceFS 100 GiB default (or streams every read from tailnet MinIO).
+
+The JuiceFS metadata + storage home is **B850** (the data-tier host). Bring it
+back first (`B850_BRINGBACK_RUNBOOK.md`) — the mounts below fail if B850 is not
+up and reachable over the tailnet.
+
+---
+
+## 0. Confirm the canonical volume BEFORE mounting
+
+There are two JuiceFS volumes and **only one is cross-node-capable**:
+
+| Volume | Storage backend | Meta | Cross-node? |
+|--------|-----------------|------|-------------|
+| **`pmoves`** | **minio** (tailnet MinIO) | Postgres (`juicefs_meta`) | ✅ **yes — mount this** |
+| `pmoves-media` | `file` (host-local disk blocks) | — | ❌ no — lists filenames, I/O-errors every read |
+
+`pmoves` is minio-backed, Postgres-meta, writable (mirror-confirmed on SPARK).
+`pmoves-media` is formatted with `Storage:"file"`, so its data blocks live on the
+formatting host's local disk and are unreachable from any other node — the
+cross-node setup script **refuses** it unless you set `ALLOW_FILE_STORAGE=1`
+(don't, for the shared mount). Verify before mounting:
+
+```bash
+make -C pmoves juicefs-storage-check
+# Expect Storage: minio (or a MinIO bucket). If it reports Storage:"file", STOP —
+# you are pointed at the wrong volume. See the storage blocker handoff.
+```
+
+---
+
+## 1. Prerequisites (per node)
+
+- Node is on the Tailscale mesh (`make -C pmoves fleet-status` shows it online —
+  the Known Road; it redacts IPs, unlike raw `tailscale status`).
+- Docker is installed and running.
+- The Supabase DB (JuiceFS meta) on B850 is reachable **by MagicDNS hostname**,
+  never a literal Tailscale/LAN IP (committed docs carry no literal IPs; the
+  DARKXSIDE egress floor fails closed on literal IPs).
+
+---
+
+## 2. Mount sequence (run on each node — 4090, then 5090)
+
+```bash
+# On the target node (4090 / 5090). JUICEFS_HOST is the MagicDNS hostname of the
+# JuiceFS meta host (B850) — use the name confirmed in B850_BRINGBACK_RUNBOOK.md.
+# DB_PASS is the Supabase DB password, sourced from the CHIT secrets pipeline
+# (exported as an env var so it reaches JuiceFS via META_PASSWORD and never
+# appears in `ps` / `docker inspect`).
+export DB_PASS=...          # from the CHIT secrets pipeline — do not paste on a shared CLI
+
+make -C pmoves juicefs-cross-node-setup \
+  JUICEFS_HOST=pmoves-b850-ai-top \
+  DB_PASS="$DB_PASS"
+```
+
+This target (`pmoves/mk/egress.mk` → `pmoves/scripts/juicefs-cross-node-setup.sh`):
+1. Pulls `juicedata/mount:ce-v1.3.0`.
+2. Runs a **storage preflight** — refuses to proceed on a `file`-backed volume.
+3. **Computes per-host bounded cache flags** via `scripts/juicefs-cache-bounds.sh`
+   (measures the `/data` volume's host backing dir) so the 4090/5090 do **not**
+   inherit the 100 GiB default and caching does not self-disable on a full disk.
+4. Mounts at `$HOME/pmoves-fs` (override with `MOUNT_POINT=`).
+
+The setup echoes the chosen cache bounds, e.g.:
+
+```
+Cache bounds: --cache-dir /data --cache-size 102400 --free-space-ratio 0.100
+```
+
+On a near-full node the helper drops `--free-space-ratio` below 0.1 (keeping
+caching enabled) and warns that the mount is too small to hold a large working
+set — heed that warning before relying on the node for heavy reads.
+
+---
+
+## 3. Verify (per node)
+
+```bash
+make -C pmoves juicefs-mount-status      # container up + content dirs visible
+ls "$HOME/pmoves-fs"                      # lists shared content
+
+# Read-path proof (the file-backed volume fails HERE with an I/O error):
+cat "$HOME/pmoves-fs"/**/<some-small-file> >/dev/null && echo "READ OK"
+
+# Cache is actually being used (not self-disabled):
+docker exec juicefs-mount sh -c 'du -sh /data/jfsCache 2>/dev/null || du -sh /var/jfsCache'
+# Expect a NON-trivial size after a few reads (bounded, not zero-because-disabled).
+```
+
+**T3 pass criteria:** 4090 and 5090 both mount the **minio-backed `pmoves`**
+volume; a read of a real file succeeds (no I/O error); `jfsCache` grows to a
+non-trivial size (cache bounded and active, not self-disabled).
+
+---
+
+## Windows note (5090)
+
+The 5090 is a Windows host. The Docker-based `juicefs-cross-node-setup` path
+assumes a Linux Docker host. For a native Windows mount, the alternative is
+**WinFsp + `juicefs.exe`** pointed at the same Postgres meta DSN (MagicDNS host,
+not an IP) with the same `--cache-dir/--cache-size/--free-space-ratio` bounds
+from `scripts/juicefs-cache-bounds.sh` (run it under Git Bash / WSL to compute the
+numbers, then pass them to `juicefs.exe mount`). Pick whichever runtime matches
+how the 5090 runs ComfyUI (WSL2 vs native).
+
+---
+
+## Related
+
+- Data-tier host bring-up → `B850_BRINGBACK_RUNBOOK.md`.
+- Storage-backend gotcha (`file` vs `minio`) →
+  `pmoves/docs/handoffs/juicefs-cross-node-storage-blocker-2026-08-04.md`.
+- Cache-bounds helper → `pmoves/scripts/juicefs-cache-bounds.sh`.