Risky behavior now covered

• LTXEulerAncestralRFScheduler — RF ancestral stepping, explicit vs auto-generated sigma schedules, duplicate-timestep index selection, and eta=0 determinism for LTX long-form video pipelines
• HeliosScheduler — default 3-stage sigma/timestep buffers and per-stage_index set_timesteps (pipeline tests only exercised stages=1)
• FlowMatchEulerDiscreteScheduler — shared flow-matching schedule/step contract used by SD3, Flux, Wan, and related pipelines (no dedicated scheduler test file existed)
• state_dict_utils — LoRA key conversion between diffusers-old, diffusers, PEFT, and UNet formats (every adapter load path)
• remote_utils — packed-latent validation, processor requirements, request payload headers, and response postprocessing without live endpoints

Test files added/updated

• tests/schedulers/test_scheduler_ltx_euler_ancestral_rf.py (new)
• tests/schedulers/test_scheduler_helios.py (new)
• tests/schedulers/test_scheduler_flow_match_euler_discrete.py (new)
• tests/others/test_state_dict_utils.py (new)
• tests/others/test_remote_utils.py (new)

Why these tests materially reduce regression risk

Scheduler schedule/step bugs silently change generation quality across many pipelines. LoRA conversion regressions break adapter loading for community checkpoints. Remote VAE helper validation gaps cause runtime failures only in hybrid-inference deployments. These unit tests are deterministic, run without GPUs or network access, and lock in the contracts that downstream pipeline tests assume but do not exercise directly.

Validation

PYTHONPATH=src python3 -m pytest -s -v \
  tests/schedulers/test_scheduler_ltx_euler_ancestral_rf.py \
  tests/schedulers/test_scheduler_helios.py \
  tests/schedulers/test_scheduler_flow_match_euler_discrete.py \
  tests/others/test_state_dict_utils.py \
  tests/others/test_remote_utils.py

All 42 tests passed locally.