# Methods

Source is feat/faces-expressive at 19dc7510389e6e567efe1c0de4213259a747ed10. References were exported with git show; runtime and unshipped compatibility code came from the audit harness. All 14 GLBs are byte-identical to their sources. Repository, shared clips, model files and other lanes were untouched.

The 390 × 844 scene uses framingTools.fitBoard, radius 1.35, top 0.33, share 0.85. Opponents use createBoard layout A; local position is actor.js youGround's x=0, z=1.88, scale 0.85. A neutral cylinder supplies table occlusion. Signature videos use the audit's fixed far seat, cropped at x=45, y=50, width=300, height=360 native pixels. Camera and card FX are unchanged.

Both 25-second sequences use seed 91109. Events occur at 6.00, 8.95, 11.90, 12.55, 15.00 and 18.00 seconds. Chrome for Testing on Mac mini M1 reports ANGLE Metal. Every explicit 1/60-second canvas sample is H.264-encoded through WebCodecs, then muxed and cut to MP4. This is an every-frame capture, not a wall-clock screencast with missing frames filled in. The slow cut repeats each source frame four times. Capture metadata and ffprobe retain frame counts, dimensions, codec and rate.

Cost uses a separate wall-clock RAF run with no table events, L1 alone, two seconds of warm-up and six measured seconds per rate. Script time includes pose, feet, face and render submission, excluding GPU completion. Calls and triangles come from renderer.info. Today excludes its initial draw and performs no waiting pose or render; its recorded script time is callback overhead. Headless Chrome's actual RAF cadence is reported, without assuming a 60 Hz display. No timing pass/fail threshold is used.

Checks sample 1,501 frames at 30-Hz idle interpolation for four table animals plus six L5 cases, ten cases and eight distinct animals. Each L5 animal runs the bidder/folder's role. Every frame is compared with Today at the same time. All skinned vertices are tested against the table cylinder. Feet use vertices with at least 50% leg influence, minimum height and near-sole centroid within 0.2% model height. Added lift or sliding over 0.2% height is flagged.

Head candidates lie inside an 80% head ellipsoid, excluding proximal appendage attachment seams. Vertical ray parity against live head-weighted triangles confirms candidates. Distal ears, arms, paws, wings and visible props are covered. This measures head-core penetration, not every possible triangle-to-triangle overlap.

Markers use each joint's furthest vertex with at least 50% influence. A jump above the source clip's own peak plus 0.2% model height is flagged; source idle advances for 25 seconds. Disabling all layers at 24.9 seconds checks permanent offsets. Continuing idle cycles are intentional. Earlier scan failures and corrective actions remain in checks.json.

Rendering used canonical Main allocation first, then canonical Mac Metal allocation when Main lacked RAM. Mac jobs hold both render locks and use slug-prefixed Dock Drive folders. Drivers close browser and server; routes delete their exact capture folders. Fedora-heavy steps use the specified memory scope, nice 19 and idle I/O. Nothing ran on omarchy. Shared preview reported available=false; static-page verification uses the canonical Chrome driver.