[ci][diagnostic]: compare GameCraft T2V with FA4 off/on on GB200

[SolitaryThinker]

Diagnostic only — do not merge.

PR #1760 full build #5043 produced one real HunyuanGameCraft T2V failure on GB200: MS-SSIM 0.7480415485 against the existing 0.93 threshold. Its automatic retry never reached checkout or tests; it exhausted the four-GPU capacity watchdog.

The canonical GameCraft GB200 reference was seeded with FASTVIDEO_FA4=1 using the older FA4 CuTe revision 82d6441e. Current main uses 14c37795 with Cutlass DSL 4.6, while the model snapshot and SSIM parameters remain stable.

This disposable branch makes the SSIM lane run only GameCraft T2V, sequentially on the same visible GB200 GPU in separate processes with FASTVIDEO_FA4=0 and FASTVIDEO_FA4=1. It logs the requested and resolved backend, torch/CUDA/GPU details, and both probe/test return codes. The lane always exits 2 so the diagnostic cannot be mistaken for a passing quality gate and does not invoke the production SSIM automatic retry.

Local validation:

• bash -n .buildkite/scripts/lanes/ssim.sh
• pytest fastvideo/tests/contract/test_ci_test_collection.py -q (13 passed)
• git diff --check
• pre-commit run --files .buildkite/scripts/lanes/ssim.sh (filename check passed; other configured hooks skip this shell path)

After the result is captured, this PR will be closed rather than merged.

[SolitaryThinker]

/test ssim

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews and 🤖 CI

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews and 🤖 CI

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=full-suite-passed
• check-success~=pre-commit

This rule is failing.

• #approved-reviews-by>=1
• check-success=full-suite-passed
• check-success~=pre-commit
• check-success=fastcheck-passed
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]