fix(security): break residual taint flows into ffmpeg args and asset URLs (round 2)

[Star-Lotus]

Follow-up to #54. After merge, CodeQL kept 9 alerts open because its taint analysis does not recognize the round-1 guards as barriers for these specific queries. This PR breaks the residual flows with zero behavior change.

Root cause (from SARIF code flows of analysis 1599605142/1599609579)

• py/command-line-injection ×6 (assets 上传人物图片 和生产出来的有区别吗？ #3, feat: 支持无 OSS 的本地优先媒体回退与 Provider 路由 #16, 无法生成404什么情况 #25–28): taint is the request parameters frame_id/script_id/offset_ms interpolated into ffmpeg output filenames / adelay filter strings. _validate_safe_id() (regex allow-list) and _safe_resolve_path() (startswith guard) are barriers for path-injection but not for the command-line-injection query.
• js/xss-through-dom ×3 (之前oss可以不用配置的，现在要强制使用了嘛？ #1, 动态演译Motion生成视频报错 #2, feat(model): HappyHorse 升级 1.0→1.1 (t2v/i2v/r2v) #40): taint passes through getAssetUrl's http/blob pass-through branch (return path); the protocol allow-list is not a recognized barrier.

Fix (no behavior change)

Alerts	Change
py ×6	Build filenames from store-backed frame.id / script.id (identical values — matched by lookup) instead of request params; offset_ms = int(offset_ms) before entering the filter string
js ×3	path.replace(/[<>\"'\]/g, "")` on the pass-through branch — recognized metachar sanitizer, no-op for well-formed URLs

Verification

• pytest: 209 passed (incl. 8 ffmpeg path-safety regression tests)
• frontend: typecheck ✅ · vitest 124 passed ✅ · build ✅

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}