refactor: centralize pinned video snapshot resolution

[atomantic]

Summary

• centralize the shared pinned Hugging Face snapshot-cache gate for Wan, FastVideo, LTX-2.x, and MiniMax H3 Ref2VA
• preserve each runtime's existing missing-revision policy and exact status, code, and message behavior
• give FastVideo its own resolved model path and verify the cached snapshot reaches --model-root
• add generateVideo() boundary coverage for FastVideo and Ref2VA cache and revision paths

generateVideo cyclomatic complexity falls from 148 (147 decision points) to 130 (129 decision points), measured with the same Babel AST traversal including nested callbacks.

Validation

• cd server && npm test -- --run services/videoGen/local.test.js — 313 passed
• cd server && npm test — 2,115 files passed, 1 skipped; 42,276 tests passed, 35 skipped
• git diff --check origin/main...HEAD
• one optional, tool-free Claude review at medium effort; findings inspected and disposition recorded, with two comment clarifications applied

Closes #6682