# Foxx and four-zombie animation polish — 2026-10-30

Foxx was the first imported rig tested and inspected, then the same system was checked on Reese, Honey Bunny and Giselle. All four use it in the game, including co-op presentation. No models were replaced and no generation credits were spent.

## What changed

- Two-bone leg IK holds the supporting ankle in world space while the other foot follows a lifted step. Stopping completes the step and settles into a narrower stance; turning takes alternating steps instead of rotating both feet across the floor.
- An idle pose sampled from each rig's existing retargeted clip blends with its moving pose. Breathing, lean and sway continue while idle. Footstep duration, lift and posture differ by character.
- A bounded pelvis adjustment keeps a trailing leg within reach. A small knee-bend reserve avoids the abrupt changes near full extension found during Honey Bunny's test.
- The supporting pose remains under the shot-reaction layer. Knockback carries the anchors with the body instead of snapping back to the raw walk pose. Existing struck-leg buckles, directional recoil and pose-continuous ragdoll death are retained.
- Attack timing and damage remain authoritative and unchanged. Remote movement drives the same local footwork; large snapshot corrections reset anchors rather than stretching legs toward their old location.

## Evidence

| Check | Result |
| --- | --- |
| [Actual four-rig footwork](../evidence/footwork-model-checks.json) | 44 pass: support placement, transition bounds, idle, turns, interrupted motion, relocation, attack support, remote movement and negative controls. |
| [Existing locomotion](../evidence/locomotion-model-checks.json) | 20 pass: pursuit, injury speed, committed attacks and evasion. |
| [Reaction and death regressions](../evidence/animation-checks.json) | 64 pass. Tests initialize the rendered stance; recovered-pose controls now share the same world location because ground-aware legs depend on position. |
| [Combat/network/handling logic](../evidence/footwork-tests.tap) | 45 pass. |
| [Four independent HTTP sessions](../evidence/coop-http-checks.json) | 13 pass locally against production Worker code and SQLite. |
| [Browser game](../evidence/footwork-browser-checks.json) | 56 pass, including actual four-rig support placement and the exterior-to-interior route. |

The instrument detects 0.20 m of unplanted foot travel in its disabled-IK negative control. In the deterministic straight-line tests, maximum planted-foot drift is below 0.004 mm and target error below 0.003 mm. These are **bone-position numerical checks**, not a claim that every visible shoe vertex stays perfectly grounded on arbitrary terrain. The largest tested per-frame joint change was 0.362 radians at 60 Hz.

![Foxx after pursuit, actual developer fixture](../evidence/foxx-footwork-review.png)

Foxx in the neutral-light developer fixture, paused after pursuit and settling. The foot-target error display is an ankle-target measurement, not an image-quality score.

![Four updated rigs inside the actual browser game](../evidence/four-footwork-gameplay.png)

Actual DGs renderer and production character models, arranged in the existing paused cast-inspection view. This is not an ordinary combat frame or concept render.

## Sources, decisions and limits

The updated game-dev-playbook and threejs-gltf-loading skills informed the workflow: preserve the existing approved cast, use one mixer per rig, inspect the imported hierarchy and test actual models. Three.js documents [animation blending](https://threejs.org/docs/pages/AnimationAction.html) and an optional [CCD IK solver](https://threejs.org/docs/pages/CCDIKSolver.html). The project uses an original bounded analytic two-bone solve for its known leg chains, without a new dependency or third-party animation download. Existing asset licenses remain in CREDITS.md.

All movement styles, blend rates, foot clearances and step trajectories are authored DGs tuning, not recovered COD values or newly captured mocap. This is a procedural animation refinement; authored turn clips, toe-roll refinement, hand/finger performance and full terrain-aware locomotion remain open. The current floor query covers the existing flat exterior and raised interior; arbitrary slopes, stairs and climbing are not supported. Internet latency, a four-human play session and device performance were not measured in this pass.

The reusable lesson and ledger entry were saved in the installed playbook. That installed directory has no Git metadata, so no upstream skill-repository push was performed.