Spatial Video Analysis

---
name: spatial-video-analysis
description: Admit explicitly authorized PNG frames and finite spatial scene estimates for the optional pinned Qwen spatial source component. Use for source-only qualification, bounded geometry preparation, source-clock provenance, or diagnosis of its provider refusals and runtime grant block. Do not use this component as proof of physical identity, motion, visibility, calibrated geometry, or an accepted spatial runtime.
---

# Spatial video analysis source component

## Installed workflow resources

For a Claude installer layout, use the adjacent managed support directory
`../video-research-resources/` and its `../video-research-resources/docs/integrations/qwen-spatial.md`.
The `../video-research-resources/integrations/qwen/video-spatio.json` records the selected component and requirements.
Run `../video-research-resources/scripts/spatial_session.py` with the explicitly selected interpreter; its adjacent helpers are installed together.
For a qualified private runtime, use `../video-research-resources/scripts/spatial_launch.py`
with a trusted interpreter and `-I -S -B` so installed bytes are admitted before
the selected interpreter starts.
The repository-relative references below apply when using a source checkout or unpacked npm package. Installing resources does not activate optional runtimes, providers or external source components.

The component is currently **runtime blocked**. The selected descriptor has
`runtime_clearance: blocked-missing-font-grant`, a null selected executable and
no bootstrap selection. The narrow scientific profile has 41 Matplotlib legacy
AFM fonts without a mapped grant; the original profile's OpenCV/FFmpeg grant
contradiction remains retained. Neither profile may activate. The frozen
28-control geometry journey with eight authorized PNGs is **UNRUN**; the owned
source/stub component is not component acceptance.

Read `docs/integrations/qwen-spatial.md` and the current
`integrations/qwen/video-spatio.json` before selecting a route. The descriptor
binds 54 external source bodies, the complete Apache-2.0 `LICENSE` and all 19
unmodified original tool specifications at the selected revision. This
repository ships first-party adapters and attribution, without copied external
source or bundled assets. Preserve original schemas and bytes.

Use `scripts/spatial_session.py` through an explicit Python interpreter with
`-I -B`. Supply `--source-root`, `--manifest`, independently trusted
`--manifest-sha256`, `--inputs`, independently trusted `--inputs-sha256`, an absent
`--output`, and `--check`. The check is read-only and stdlib-only. A source/input
admitted but runtime-blocked JSON result exits 2, creates no output and imports
no foreign source or scientific package. Serving also refuses before foreign
import. Do not treat a successful source readback as runtime clearance, create
a substitute profile or run the unaccepted geometry journey.

Input authority selects absolute regular PNG paths with exact SHA256/byte
counts, original source frame IDs and actual PTS/time-base clocks. Optional
scene files need their own path/hash/byte selection. Keep at most 64 frames and
16 million pixels per PNG, strictly increasing clocks and unique paths/IDs.
Only selected image/scene paths are permitted; legacy NPY/base64 branches are
outside this route. Validate scene dimensions, finite cameras/intrinsics,
positive depth/scale, ordered contained boxes and matching frame keys before
dispatch. Reject duplicate IDs within a frame and conflicting ID labels.

All actual VLM dispatch is denied, including errors swallowed by upstream
experts. Keep orientation/grounding tools advertised with
`provider_unavailable`; never turn refusal into absence, visibility or grounding
evidence. Deterministic viewpoint layout delegates the original expert without
images and labels assumed facing/model orientation unverified. Refuse missing
frames, missing or ambiguous viewpoint targets and unknown operations. Use
the owned corrected mobile signatures rather than the original broken dispatch.

Preserve generated frame-qualified IDs and label cross-view correspondence
heuristic. Caller-supplied persistent IDs are unverified assertions. Preserve
two same-frame instances; refuse proximity components that could merge them.
Map uniform keyframe positions back to source IDs. Missing object motion is
unknown, with no stationary claim. Owned Pillow appearance segmentation uses
float32 differences, explicit Pillow BILINEAR resampling, the documented strict
threshold/minimum-run contract and source IDs; declare its difference from
OpenCV and its lack of physical-motion proof. Point extrapolation delegates the
original expert with finite 2D/3D inputs, complete outputs, at most 32 future
steps and order 5.

Carry source clock/revision provenance and depth, camera, assumed FOV60 optics,
identity and metric-scale caveats beside every output. Future eligible sessions
must use exactly selected runtime metadata/bootstrap bytes, owned cwd/config/
cache/MPL/temp directories, credential whitelist and actual loaded-file hash
receipts. The concrete launcher requires a standard-library venv with zero
startup hooks and a complete installed byte inventory. Its Agg font boundary
admits only 38 bundled TTFs and records font requests separately from module
imports. This process has host privileges and is not a sandbox. Keep source,
owned unit checks, runtime grants and physical/model acceptance as separate
claims. Scoped first-party motion units using previously qualified root
NumPy/Pillow remain unit evidence, with no selected-runtime qualification.