test(e2e): deterministic H3 t2va 2-GPU flowGRPO metric to prevent regression

[Rockdu]

What

• --deterministic-mode on the H3 t2va 2-GPU flowGRPO recipe, matching the SD3 recipes that back a metric-regression e2e.
• A new tests/e2e/short/test_h3_t2va_grpo_2xGPU.py (stage-c-3-gpu-h200, 2 rollouts) with its recorded standard, so H3 gets the same bit-for-bit metric regression the other families already have.

CI sizing: the test runs the recipe at --n-samples-per-prompt 4 --eval-interval 0 (8 videos per rollout, no eval). The recipe's own sizes put one tp2 engine at ~60 min for two rollouts, past the suite's timeout; this config keeps both rollouts — so the LoRA-IPC weight sync between them stays covered — at 1492 s in CI.

Why

H3 is the family with the most model-specific rollout code (its own packed-sequence denoise loop, its own SDE update, its own LoRA-IPC weight key mapper), and no e2e guarded it — regressions there only surfaced on the 17/34-GPU production runs.

Validation

Recorded on H200 against main before #215/#216, with stock sglang-miles-h3, --deterministic-mode, 2 rollouts:

Metric	rollout 0	rollout 1
rollout/reward/raw_mean	0.7477782	0.7937961
rollout/reward/raw_std	0.0313390	0.0288071
train/log_prob_mean_abs_diff	5.75e-05	4.13e-05

log_prob_mean_abs_diff lands in the e-5 band the H3 recipe is verified against, i.e. the train forward reproduces the rollout trajectory; reward moves in the expected direction across the two rollouts.

The same recipe was also recorded on top of #216 (SDE-window trajectory transport) with its sglang counterpart: all eight metric series came out bitwise identical to this baseline, which is the H3-path evidence that the windowed transport changes nothing observable. CI on this test then passed against the standard, so the run is reproducible across machines, not just locally.

Files

• scripts/run_diffusion_grpo_h3_t2va_2gpu.py — --deterministic-mode.
• tests/e2e/short/test_h3_t2va_grpo_2xGPU.py + tests/ci/fixtures/e2e_standards/test_h3_t2va_grpo_2xGPU.json — the test and its standard.

Checklist

• pre-commit run --all-files passes
• Added/updated tests for new behaviour — this PR is the test
• pytest -x is green — tests/fast, 239 passed
• If launch flags changed, python3 train_diffusion.py --help still parses — no new flags
• If a public flag was added, it appears in the CLI reference docs — n/a
• If an example was added, it has a real walkthrough — n/a