fix: validate configured Fast video model input

[ch-z-hc]

Bug

After setting default_video_model to MiniMax-Hailuo-2.3-Fast, mmx video generate --prompt "A cat" submits a task without an image. The existing image requirement checks only the explicit --model flag, so it is bypassed when the same model comes from config.

Fix

Apply the existing guard to the resolved model. One product line changes; model precedence and automatic SEF/S2V selection stay unchanged.

The regression executes the command against a local HTTP fixture: original code submits a task; patched code returns usage error 2 with zero task submissions. Five controls cover --image, the hidden --first-frame alias, explicit model override, and SEF/S2V switching. A separate real CLI dry-run reads a temporary config file: original exits 0 with an invalid request preview, patched exits 2 with the existing image-required error.

This prevents task submission, not every network operation during CLI startup: existing authentication, region detection and update checks are outside this patch.

Validation

• Original command tests with regressions: 19 passed, 1 failed. Patched command + video SDK tests: 36 passed on Windows and Linux.
• Windows/Linux typecheck passes; lint has no errors and one existing speech.test.ts warning. Linux production build and compiled CLI version check pass.
• Two independent automated blind reviews found no actionable issues. One independently reproduced original task submission and patched rejection, and ran a 72-test affected selection plus seven additional precedence/alias cases.
• Full Linux run: 602 passed, 3 failed. Unchanged a965e72 in the same environment: 596 passed, the same 3 failed (unreachable-region timeouts in timeout-fix.test.ts and auth/login.test.ts, and the FileSDK upload base-URL assertion). Their causes were not diagnosed here. Both use the same validation-only fetch preload rejecting non-loopback destinations by default; test-owned mocks can override it. No repository assertions, timeouts or skip conditions were relaxed.

No live MiniMax model calls. Implemented and tested with Codex assistance; the independent blind reviews were automated, not human approval.

---

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