fix: enforce H3 reference input limit

[shaoohh]

Summary

• enforce MiniMax-H3's combined limit of 12 reference images, videos, and audios
• keep the existing per-media limits of 9 images, 3 videos, and 3 audios
• add regression coverage proving 12 mixed references are accepted and 13 are rejected

Root cause

The H3 validator checked each media type independently but did not check their combined count. A request containing 9 reference images, 3 reference videos, and 3 reference audios therefore passed local validation with 15 total references, even though mixed-reference requests allow at most 12.

Impact

Oversized mixed-reference requests now fail locally with an actionable usage error before a paid generation request is submitted.

Validation

• bun test test/video/v2.test.ts — 12 passed
• bun test — 450 passed
• bun run typecheck
• bun run lint — 0 errors; 1 pre-existing warning in test/sdk/speech.test.ts
• bun run build

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[NianJiuZst]

Could you share the authoritative source for the 12-item combined reference limit? The current Video Generation V2 documentation describes reference-to-video as accepting any combination of reference images, videos, and audio, and documents only the per-type caps (9 images, 3 videos, and 3 audios), without a combined cap. With this change, a request containing 9 images, 3 videos, and 1 audio is rejected locally even though it is within those published per-type limits.

The implementation and tests look correct if the backend really enforces 12 total. Before merging, please either link the relevant service/spec evidence or update the public docs so the CLI does not risk rejecting a documented-valid request.