fix(hunyuan-video15): align vLLM-Omni rollout policy

[leviking98z-rgb]

Summary

Align HunyuanVideo-1.5 trainside and vLLM-Omni at the input, stochastic
transition, checkpoint, and online-LoRA boundaries:

• preserve the exact upstream MLLM system-prompt whitespace because
  mllm_crop_start=108 is tied to that tokenized prefix;
• keep fp32 LoRA master parameters on both matched recipes while retaining
  BF16 forward parameters;
• derive SDE noise from the same (seed, step, sample_id) tuple on both
  engines, preserving per-sample independence across vLLM request splitting;
• form the transformer timestep as fp32 sigma * 1000 before the BF16 cast,
  matching Diffusers/vLLM;
• select the trainer's SDPA backend through the small boot-intent path used by
  the pinned vLLM-Omni selector;
• load the token refiner's unfused Q/K/V checkpoint tensors instead of leaving
  them randomly initialized;
• expose the main blocks' packed QKV LoRA mapping and apply all 12 configured
  token-refiner nn.Linear LoRA targets, failing fast on incomplete coverage;
  and
• keep algorithm.old_logp_source: rollout, using the sampler-emitted density
  as the frozen policy anchor.

The cleanup deliberately excludes the pin-specific exact-arithmetic
monkey-patches for autocast, Q/K RMSNorm, SDPA mask construction, and RoPE, as
well as the rejected merged-weight and raw-VAE-reward experiments.

Related Issue

Related to #386.

Test Plan

• python -m py_compile on all changed Python modules — passed.
• ruff check and ruff format --check on all changed Python modules —
  passed.
• pre-commit run --files <all changed files> — all applicable hooks passed,
  including YAML, AST, Ruff, formatting, recipe target, docstring,
  experimental-boundary, and dependency-direction checks.
• Static AST comparison confirms the trainside prompt template is byte-equal
  to both pinned Diffusers and vLLM-Omni (477 characters).
• YAML assertions confirm both recipes resolve master_dtype: fp32, the
  vLLM-Omni recipe resolves old_logp_source: rollout, and the stage resolves
  attention_backend: TORCH_SDPA.
• The prompt-only matched run changed first-rollout reward delta from the
  unpatched +0.0370 to -0.0007, confirming that prompt whitespace is an
  initial-conditioning requirement rather than a cosmetic edit.
• The scheduler contract produced bit-identical emitted transitions and log
  probabilities on both engines while assigning different SDE noise to
  different samples.
• The timestep contract found the old trainside ordering produced 976
  versus 980 and 684 versus 680 at two steps of the matched schedule;
  fp32 scaling makes all ten transformer timesteps bit-identical.
• A full-checkpoint audit found 12 unequal refiner Q/K/V parameters before the
  loader fix. After it, all 15.510 GiB compared bit-identically
  (missing=0, unequal=0).
• Live LoRA coverage increased from 324 to 444 active wrappers. In the
  loader-only versus full-coverage comparison at rollout 75, correlation
  improved from 0.964188 to 0.992005 and MAE from 0.00398356 to
  0.00226971.
• The accepted full-coverage arm remained aligned past the suspected failure
  point: correlation/MAE were 0.999703 / 0.00046770 through rollout 10 and
  0.991211 / 0.00354971 through rollout 125. At rollout 140 correlation was
  0.990625 with mean bias +0.00309; no ratio or reward-curve collapse
  occurred:
  • vLLM-Omni run
  • trainside reference

That live curve included the exact-arithmetic probes now excluded from the
final diff. The selected input/weight/LoRA contracts were validated within
that run, but the narrowed final HEAD has not been rerun on GPU; the PR
therefore remains Draft.

Compatibility / Risk

• The runtime LoRA/checkpoint patches are gated to the pinned
  HunyuanVideo-1.5 vLLM-Omni model. Existing packed checkpoint loading for the
  54 main blocks remains unchanged.
• The sample-keyed SDE path is opt-in: it is used only when the request carries
  denoise_seeds; existing callers without those keys keep their prior RNG
  behavior.
• Token-refiner adapters use two ordinary BF16 linear operations instead of a
  Punica wrapper, adding a small refiner-forward cost.
• The targeted refiner wrapper supports adapter slot 0, matching the recipe's
  single-online-adapter setup, and rejects an unexpected slot explicitly.
• No checkpoint schema, dataset format, or trajectory wire-format migration is
  introduced.

Reviewer Notes

The exploratory history is intentionally retained. Commit 89e29069 removes
the broad candidate set additively; commit c95c7768 restores only the
validated input/training contracts and small backend selection path. The final
runtime contains exactly three HunyuanVideo-1.5 monkey-patches, all for
checkpoint or LoRA coverage.

Duplicate-work check: #220 covers HunyuanVideo-1.0 conditioning/merged rollout
weights, while #371 covers the SGLang pipeline/SDE path and supports both
sample-keyed noise and the rollout-logp anchor choice. Neither applies all
HunyuanVideo-1.5 vLLM-Omni packed/refiner LoRA targets. AI assistance was used;
the final diff and reported validation results were reviewed before
publication.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.

EXP