[codex] Improve hatch-pet playback QA

[sailxjx]

Summary

• Add hatch-pet guidance for playback-specific QA issues: per-frame size popping, clipped wide poses, wrong facing direction, non-alternating gait, near-identical idle frames, and hidden transparent-pixel color residue.
• Add fixed-slot frame extraction support to preserve row-level scale and placement when generated strips already have stable character sizing.
• Preserve frame order when deriving running-left from running-right, and validate the per-frame mirror in finalization.
• Clear hidden RGB values in fully transparent atlas pixels and save WebP with exact alpha preservation.

Why

While building and previewing custom pets, several issues passed geometry validation but showed up during animation playback: cropped frames made characters pop in size, whole-strip mirroring reversed timing, directional rows could face the wrong way, and transparent RGB residue could appear as colored blocks in some viewers. These changes make the deterministic pipeline and skill instructions capture those failure modes directly.

Validation

• python3 -m py_compile on the modified hatch-pet scripts.
• git diff --check.
• Ran script --help checks with the Codex bundled Python runtime.
• Ran a small functional check for per-frame mirroring, transparent RGB cleanup, and fixed-slot extraction.