diff --git a/backend/app/main.py b/backend/app/main.py
index 84599ac..9c79ce1 100644
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -698,7 +698,11 @@ async def get_asset_content(asset_id: str) -> FileResponse:
     abs_path = asset_store.get_absolute_path(rec)
     if not abs_path.is_file():
         raise HTTPException(status_code=404, detail="Asset file missing on disk")
-    return FileResponse(abs_path, media_type="image/png")
+
+    import mimetypes
+    guessed_type, _ = mimetypes.guess_type(str(abs_path))
+    content_type = guessed_type or ("video/mp4" if rec.media_type == "video" else "image/png")
+    return FileResponse(abs_path, media_type=content_type)
 
 
 @app.post("/api/v1/workflow/cancel/{run_id}")
diff --git a/docs/ACCEPTANCE_DOSSIER.md b/docs/ACCEPTANCE_DOSSIER.md
index 8a2c37b..b337b78 100644
--- a/docs/ACCEPTANCE_DOSSIER.md
+++ b/docs/ACCEPTANCE_DOSSIER.md
@@ -47,25 +47,35 @@ All changes have been developed on focused feature branches and integrated into
 | **[PR #14](https://github.com/BerryUIKI/AI-Studio/pull/14)** | `feature/m10-workflow-repair` | `feat(comfy): workflow validation and automated repair engine (M10)` | 3/3 Passed | [`60c3620`](https://github.com/BerryUIKI/AI-Studio/commit/60c3620) |
 | **[PR #15](https://github.com/BerryUIKI/AI-Studio/pull/15)** | `feature/m11-discrete-gpus` | `feat(hardware): support non-NVIDIA discrete GPUs with DirectML and launch flag injection (M11)` | 3/3 Passed | [`7f700b2`](https://github.com/BerryUIKI/AI-Studio/commit/7f700b2) |
 | **[PR #16](https://github.com/BerryUIKI/AI-Studio/pull/16)** | `feature/m12-cross-platform` | `feat(platform): cross-platform desktop adapters for macOS Metal and Linux (M12)` | 3/3 Passed | [`2551796`](https://github.com/BerryUIKI/AI-Studio/commit/2551796) |
+| **[PR #17](https://github.com/BerryUIKI/AI-Studio/pull/17)** | `docs/dossier-roadmap-sync` | `docs: synchronize acceptance dossier and roadmap post v0.1 with dev integration commits` | 3/3 Passed | [`0a59aba`](https://github.com/BerryUIKI/AI-Studio/commit/0a59aba) |
+| **[PR #18](https://github.com/BerryUIKI/AI-Studio/pull/18)** | `feature/repair-agent-safety` | `fix(workflow): enforce model family guards and ambiguity blocking in workflow repair` | 3/3 Passed | [`3cf860a`](https://github.com/BerryUIKI/AI-Studio/commit/3cf860a) |
+| **[PR #19](https://github.com/BerryUIKI/AI-Studio/pull/19)** | `feature/hardware-matrix-realism` | `fix(hardware): accurately classify AMD, Intel, and Apple Silicon as candidate or source-compatible` | 3/3 Passed | [`c35f3b1`](https://github.com/BerryUIKI/AI-Studio/commit/c35f3b1) |
+| **[PR #20](https://github.com/BerryUIKI/AI-Studio/pull/20)** | `feature/video-format-dossier-alignment` | `fix(video): support video containers and animated webp, fix MIME type, align dossier evidence` | 3/3 Passed | Pending |
 
 ---
 
-## 3. Supported Versions & Hardware
+## 3. Supported Versions & Hardware Evidence Tiers
 
-### Operating Systems
-- **Windows 10 / 11 (64-bit)**: Primary target platform for portable release distribution.
-- **Cross-Platform Core**: macOS / Linux supported for source checkouts; verified through core portable test suite.
+### Operating Systems & Platform Verification Status
+- **Windows 10 / 11 (64-bit)**: **Physically Verified**. Tested on Windows 11 host with native Rust launcher, ComfyUI engine supervisor, single-instance Win32 mutex, and local inference.
+- **macOS (Apple Silicon, Metal MPS)**: **Source Compatible**. Directory paths, browser launcher commands, Unix domain socket single-instance checks, and Metal fp16 flags implemented and unit-tested; native macOS binary bundling and physical Apple Silicon execution remain unverified in this session.
+- **Linux (Ubuntu 22.04+)**: **Source Compatible**. Supervisor process handling, Unix domain socket locking, and headless CLI generation implemented and unit-tested; native Linux binary packaging and physical non-NVIDIA execution remain unverified in this session.
 
-### Hardware Tiers
+### Hardware Tiers & Real Evidence Separation
 - **Tier 1 (Cloud-Only / Featherweight)**:
   - **GPU**: None required (Zero GPU, Zero PyTorch on host).
   - **RAM**: 4 GB RAM minimum.
   - **Storage**: < 500 MB disk footprint.
-  - **Mode**: BYOK Cloud APIs (OpenAI, Fal.ai, SiliconFlow).
-- **Tier 2 (NVIDIA Local Inference)**:
-  - **GPU**: NVIDIA RTX (Turing, Ampere, Ada Lovelace, Blackwell) with minimum 6 GB VRAM (8 GB+ recommended).
-  - **Driver**: NVIDIA Driver 535+ with CUDA 12.1+.
-  - **Engines**: ComfyUI (managed or external), Stable Diffusion WebUI (AUTOMATIC1111).
+  - **Mode**: BYOK Cloud APIs (OpenAI, Fal.ai, SiliconFlow). Real API calls verified.
+- **Tier 2 (Physically Verified NVIDIA CUDA)**:
+  - **GPU**: NVIDIA RTX (RTX 3060 12GB Ampere, driver 572.70, CUDA 12.8 physically verified on host rig).
+  - **Engines**: ComfyUI (managed and external), Stable Diffusion WebUI (AUTOMATIC1111).
+- **Tier 3 (Candidate Architecture - Unverified Physical Non-NVIDIA Discrete GPUs)**:
+  - **Hardware**: AMD Radeon (DirectML / ROCm) and Intel Arc (DirectML / OneAPI IPEX).
+  - **Status**: Detection and launch flag injection implemented and verified via unit tests; physical generation unverified in this session.
+- **Tier 4 (Source-Compatible Desktop Adapters - Unverified Native Build)**:
+  - **Hardware**: Apple Silicon (MPS Metal).
+  - **Status**: Platform adapters implemented and unit-tested; native binary compilation and physical Mac runs unverified in this session.
 
 ---
 
@@ -138,6 +148,18 @@ cargo run --manifest-path launcher/Cargo.toml
 3. **Task Cancellation**:
    - Local engine cancellation dispatches interrupt signals (`/interrupt` on ComfyUI, `/sdapi/v1/interrupt` on SD WebUI).
    - Cloud task cancellation immediately flags the task as cancelled locally, but truthfully discloses: `"Cloud cancellation requested locally. Note: external cloud providers may continue asynchronous inference or incur compute charges."`
+4. **Video Generation & Container Formats**:
+   - Fal.ai (Fast SVD) and SiliconFlow (CogVideoX) output native MP4 (`video/mp4`, H.264).
+   - Standard ComfyUI (AnimateDiff/SVD) without custom nodes outputs animated WebP (`image/webp`) via `SaveAnimatedWEBP`. If `VHS_VideoCombine` with ffmpeg is installed, ComfyUI outputs native MP4.
+   - In the frontend canvas, MP4 and WebM containers render natively in `<video controls loop>`, while animated WebP renders via `<img>` with an "Animated WebP" badge.
+   - Downloads and exports strictly preserve the native container extension (`.mp4`, `.webm`, or `.webp`).
+5. **Conversational Agent Boundaries**:
+   - The assistant operates via deterministic keyword and regular expression intent matching against predefined creative action schemas.
+   - It does not claim general unconstrained natural-language reasoning or autonomous arbitrary workflow synthesis.
+   - All execution requires explicit user review and approval via the proposal confirmation gate.
+6. **Workflow Repair Ambiguity & Compatibility Guards**:
+   - The ComfyUI repair engine automatically reconnects slots only when an unambiguous single source exists in the DAG. If multiple sources exist, automatic connection is blocked (`substitution_blocked`) to avoid graph corruption.
+   - Checkpoint substitution is strictly constrained to matching architecture families (`sd15`, `sdxl`, `flux`, `svd`). Unknown or cross-family substitutions are blocked.
 
 ---
 
diff --git a/docs/ROADMAP_POST_V0_1.md b/docs/ROADMAP_POST_V0_1.md
index 8f9eb5d..b3a8a2b 100644
--- a/docs/ROADMAP_POST_V0_1.md
+++ b/docs/ROADMAP_POST_V0_1.md
@@ -53,7 +53,9 @@ This roadmap operationalizes the long-term approved technical directions for Ber
     - AnimateDiff SD1.5 / SDXL macro subgraphs (`AnimateDiffLoaderWithContext`, `KSampler`, `VHS_VideoCombine` / `SaveAnimatedWEBP`).
 - **Technical Boundaries**:
   - Video port type conforms to the 5-type contract (`video`).
-  - Output stored in `AssetStore` with media type `video/mp4` or `video/webm`, content-addressed via SHA-256.
+  - Output stored in `AssetStore` with media type `video/mp4`, `video/webm`, or `image/webp` (for animated WebP from standard ComfyUI). Content-addressed via SHA-256.
+  - Video playback in GUI: native `<video>` for MP4/WebM containers; `<img>` with "Animated WebP" badge for animated WebP outputs.
+  - Export: downloads preserve actual container extension (`.mp4`, `.webm`, or `.webp`).
   - Deterministic dirty-check caching accounts for video-specific parameters: `motion_bucket_id`, `fps`, `num_frames`, `duration`.
 - **Data & Project Compatibility**:
   - SQLite database persists video assets and execution records. Project schema v1 loads seamlessly; canvas renders video cards.
@@ -107,11 +109,12 @@ This roadmap operationalizes the long-term approved technical directions for Ber
   - The UI presents the plan with an **"Approve & Run"** button and editable fields.
   - The user can adjust parameters or approve execution; the Agent dispatches the action to `CreativeRunner` and presents results with next-step suggestions.
 - **Supported Capability Matrix**:
-  - Natural-language mapping to all creative actions (txt2img, img2img, inpaint, upscale, txt2video).
-  - Intent classification, parameter extraction, and template matching.
+  - Deterministic keyword and regex mapping to creative actions (txt2img, img2img, inpaint, upscale, txt2video, img2video).
+  - Parameter extraction (aspect ratios, step count, seeds, engines) and template recommendation.
 - **Technical Boundaries**:
-  - Backend `AgentService` with rule-based and LLM-assisted intent resolution.
+  - Backend `AgentService` with rule-based and regex intent resolution; does not claim general unconstrained natural-language understanding or arbitrary workflow synthesis.
   - Strict human-in-the-loop gate: execution requires user approval token.
+  - Transparent plan disclosure with cost, engine, and parameter visibility.
   - Never mutates engine files or executes uninspected external commands.
 - **Data & Project Compatibility**:
   - Conversation sessions and generated proposals are persisted in project SQLite tables.
@@ -120,35 +123,37 @@ This roadmap operationalizes the long-term approved technical directions for Ber
   - UI displays proposal, parameters, and confirmation dialog.
   - Execution only dispatches upon explicit user confirmation.
 - **Explicit Exclusions**:
-  - Autonomous loop execution without user confirmation; silent billing against cloud accounts.
+  - Unbounded conversational reasoning without explicit action templates; autonomous loop execution without user confirmation; silent billing against cloud accounts.
 
 ---
 
 ### M10: Agent-Assisted Construction, Validation, Repair, and Execution of ComfyUI Workflows
 
 - **User Journey**:
-  - Users can import complex ComfyUI workflows or ask the Agent to *"Add face detailing to my workflow"* or *"Diagnose why this workflow fails"*.
+  - Users can import complex ComfyUI workflows or ask the Agent to *"Diagnose why this workflow fails"*.
   - The Agent analyzes the ComfyUI DAG:
     - Validates node connections against port types (`MODEL`, `CLIP`, `VAE`, `LATENT`, `IMAGE`).
     - Checks for missing custom node classes and missing checkpoint/LoRA models.
     - Identifies topological cycles or disconnected subgraphs.
-  - If errors exist, the Agent proposes a concrete **Repair Diff** (e.g. *"Node 8 (VAEDecode) is missing VAE input; connect from Node 4 [slot 2]"*).
+  - If errors exist, the Agent proposes a concrete **Repair Diff** (e.g. *"Node 8 (VAEDecode) is missing VAE input; connect from unambiguous Node 4 [slot 2]"*).
+  - **Ambiguity Guards**: When multiple potential sources exist for an input slot (e.g., multiple VAE loaders or EmptyLatentImage nodes), automatic reconnection is blocked (`substitution_blocked`) rather than guessing, preventing corrupted dataflow.
+  - **Model Family Guards**: Missing checkpoints are substituted only within the exact same architecture family (`sd15`, `sdxl`, `flux`, `svd`); cross-family or unknown substitutions are strictly blocked.
   - The user inspects the visual diff and accepts the repair.
   - Once validated, the workflow can be executed directly or registered into the `TemplateRegistry`.
 - **Supported Capability Matrix**:
-  - ComfyUI DAG syntax parser, topological sort validator, and port type-checker.
-  - Common pattern repairs: missing VAE decode, mismatched latent dimensions, missing LoRA bypass.
-  - Model dependency resolver: checks against indexed local models.
+  - ComfyUI DAG syntax parser, topological sort cycle detector, and port type-checker.
+  - Safe pattern repairs: unambiguous single-source slot reconnections, same-family model substitution.
+  - Model dependency resolver: checks against indexed local models with family inference.
 - **Technical Boundaries**:
   - Pure Python DAG analysis engine in `backend/app/core/workflow_validator.py` and `backend/app/core/workflow_repair.py`.
   - Sandboxed validation without running untrusted Python node code.
 - **Data & Project Compatibility**:
   - Validated workflows can be saved as custom templates in `TemplateRegistry`.
 - **Acceptance Criteria**:
-  - Automated tests verify detection of broken connections, missing models, and invalid node types.
-  - Repair engine fixes broken workflows and produces valid, executable graphs.
+  - Automated tests verify detection of broken connections, missing models, cycles, and type mismatches.
+  - Repair engine fixes broken workflows safely and blocks ambiguous or cross-family mutations.
 - **Explicit Exclusions**:
-  - Arbitrary Python code execution inside custom nodes; automated downloading of untrusted git repositories.
+  - Arbitrary Python code execution inside custom nodes; automated downloading of untrusted git repositories; guessing ambiguous wiring.
 
 ---
 
@@ -158,12 +163,12 @@ This roadmap operationalizes the long-term approved technical directions for Ber
   - Users with AMD Radeon or Intel Arc discrete GPUs run Berry AI Studio.
   - The Environment Manager detects the GPU hardware, reports detected VRAM, and displays the appropriate acceleration backend (DirectML / ROCm / OneAPI IPEX).
   - Engine supervisors automatically pass vendor-optimized launch flags to ComfyUI (e.g., `--directml`, `--use-split-cross-attention`, `--lowvram`).
-  - Unsupported or untested configurations display clear performance caveats and recommendation to use Cloud BYOK if local performance is inadequate.
+  - Candidate configurations are clearly labeled as "Candidate Architecture (Unverified on Physical Hardware in Current Session)" with guidance to use Cloud BYOK if local performance is inadequate.
 - **Supported Capability Matrix**:
-  - *Reference Hardware*:
+  - *Candidate Reference Hardware*:
     - AMD Radeon RX 7900 XTX / 7800 XT / 6700 XT (Windows DirectML, Linux ROCm 6.x).
     - Intel Arc A770 / A750 (Windows DirectML, Linux IPEX/OneAPI).
-  - *Status Classification*: Explicitly distinguished as "Verified on Reference Hardware" vs. "Experimental / DirectML Emulated".
+  - *Status Classification*: Explicitly distinguished as "Verified on Physical Hardware" (NVIDIA CUDA) vs. "Candidate Architecture (Unverified on Physical Hardware in Current Session)" vs. "Experimental / DirectML Emulated".
 - **Technical Boundaries**:
   - Extend `backend/app/runtime/hardware.py` to detect AMD GPUs via `rocm-smi` (Linux) and WMI `Win32_VideoController` (Windows), and Intel Arc GPUs via WMI.
   - Update `HardwareReadiness` schema with vendor classification (`nvidia`, `amd`, `intel`, `apple_silicon`, `cpu_only`).
@@ -173,9 +178,9 @@ This roadmap operationalizes the long-term approved technical directions for Ber
 - **Acceptance Criteria**:
   - Hardware detector correctly identifies AMD and Intel GPUs and reports VRAM.
   - Engine supervisor injects `--directml` or vendor flags when non-NVIDIA GPU is selected.
-  - `docs/SUPPORT_MATRIX.md` documents verified vs. unverified combinations.
+  - `docs/SUPPORT_MATRIX.md` documents verified vs. candidate and unverified combinations without claiming physical verification for unverified hardware.
 - **Explicit Exclusions**:
-  - Custom kernel compilation; driver installation scripts; obsolete legacy GPUs lacking FP16 support.
+  - Custom kernel compilation; driver installation scripts; claiming physical hardware verification without physical generation run logs.
 
 ---
 
diff --git a/frontend/src/components/canvas/ImageCardNode.tsx b/frontend/src/components/canvas/ImageCardNode.tsx
index 3dd5c7a..2292f39 100644
--- a/frontend/src/components/canvas/ImageCardNode.tsx
+++ b/frontend/src/components/canvas/ImageCardNode.tsx
@@ -11,12 +11,18 @@ export const ImageCardNode = memo(({ id, data, selected }: NodeProps) => {
   const { nodes } = useCanvasStore();
 
   const isVideo = cardData.mediaType === 'video' || Boolean(cardData.videoUrl);
+  const targetMediaUrl = cardData.videoUrl || cardData.imageUrl || '';
+  const isVideoContainer = targetMediaUrl.toLowerCase().match(/\.(mp4|webm|mov)(\?|#|$)/) !== null;
 
   const handleExport = (e: React.MouseEvent) => {
     e.stopPropagation();
     const link = document.createElement('a');
-    link.href = cardData.videoUrl || cardData.imageUrl;
-    const ext = isVideo ? 'mp4' : 'png';
+    link.href = targetMediaUrl;
+    let ext = isVideo ? (isVideoContainer ? 'mp4' : 'webp') : 'png';
+    const match = targetMediaUrl.match(/\.([a-zA-Z0-9]+)(?:\?|#|$)/);
+    if (match && ['mp4', 'webm', 'mov', 'png', 'jpg', 'jpeg', 'webp', 'gif'].includes(match[1].toLowerCase())) {
+      ext = match[1].toLowerCase();
+    }
     link.download = `${cardData.label || 'berry_asset'}.${ext}`;
     document.body.appendChild(link);
     link.click();
@@ -111,9 +117,9 @@ export const ImageCardNode = memo(({ id, data, selected }: NodeProps) => {
 
       {/* Media Preview Container */}
       <div className="relative bg-slate-950 flex items-center justify-center min-h-[220px] max-h-[360px] overflow-hidden">
-        {isVideo ? (
+        {isVideo && (isVideoContainer || !targetMediaUrl.toLowerCase().match(/\.(webp|gif)(\?|#|$)/)) ? (
           <video
-            src={cardData.videoUrl || cardData.imageUrl}
+            src={targetMediaUrl}
             controls
             loop
             playsInline
@@ -121,12 +127,17 @@ export const ImageCardNode = memo(({ id, data, selected }: NodeProps) => {
           />
         ) : (
           <img
-            src={cardData.imageUrl}
+            src={targetMediaUrl || cardData.imageUrl}
             alt={cardData.label || 'Generated creative asset'}
-            className="w-full h-auto object-contain select-none pointer-events-none"
+            className="w-full h-auto object-contain select-none"
             loading="lazy"
           />
         )}
+        {isVideo && !isVideoContainer && (
+          <div className="absolute top-2 left-2 px-2 py-0.5 rounded-md bg-purple-950/80 backdrop-blur text-[10px] font-medium text-purple-300 border border-purple-800/60">
+            Animated WebP
+          </div>
+        )}
         {p && (
           <div className="absolute bottom-2 left-2 px-2 py-0.5 rounded-md bg-slate-950/75 backdrop-blur text-[10px] font-mono text-slate-300 border border-slate-800">
             {p.dimensions}