chore(scripts): drop ad-hoc verify harnesses, move dataset converters to datasets/ - #284

Summary

Repo policy since #99/#267 is no unenforced test tree — yet scripts/ was quietly regrowing one: #158 left seven ep_verify/ torchrun harnesses and #210 a KL-batching regression script. This PR:

• Deletes scripts/ep_verify/ (7 files) and scripts/verify_refl_kl_batching.py. They were verification evidence for their PRs; they remain recoverable from git history (feat(veomni-ep): expert parallelism for the VeOmni FSDP2 training backend #158 / feat(experimental): add WAN ReFL as a self-contained training package #210) if anyone needs to re-run them.

• Moves the two dataset converters next to the data they produce, following the existing datasets/<name>/ convention (cf. datasets/pickscore/prpocess.py):

  • scripts/convert_video_r1_260k_to_unirl.py → datasets/video_r1_260k/
  • scripts/convert_daily_omni_dataset_format_to_unirl.py → datasets/daily_omni_av/

  The qwen3-omni recipes' default data_path already points into exactly these directories, so converter output now lands beside the converter.

• Fixes a copy-paste bug in the daily-omni converter docstring (its example invoked the video-r1 script) and updates self-referencing paths.

scripts/ now holds only check_recipe_targets.py — the pre-commit/CI guard, untouched here (#283 is reworking it separately).

Related Issue

N/A

Test Plan

• python scripts/check_recipe_targets.py → check-recipe-targets: 2310 unirl _target_ paths resolve.
• python -m py_compile on both moved converters → OK
• ruff check + ruff format --check on the moved files → clean
• git grep "scripts/convert_\|scripts/ep_verify\|scripts/verify_refl" → no references anywhere (the removed/moved files were referenced nowhere outside scripts/; no CI workflow touches them)

Compatibility / Risk

• No shipped-code changes: neither scripts/ nor datasets/ is packaged (wheel includes unirl.* only).
• ⚠️ Merge-order interaction: feat(experimental): tier contract, boundary lint, private-package convention #279 and docs(algorithms): placement policy — core canon vs experimental variants (RFC) #280 branched before feat(experimental): add WAN ReFL as a self-contained training package #210 merged and still carry scripts/verify_refl_kl_batching.py as an addition — merging them after this PR without a rebase would resurrect the file. Please drop it from those branches when rebasing.
• perf(lint): make the recipe _target_ guard filesystem-cheap #283 only touches scripts/check_recipe_targets.py; no overlap.

Reviewer Notes

AI-assisted (Claude Code); the submitter reviewed the full diff. Duplicate-work check: scanned all open PRs touching scripts/ (#279, #280, #283) — none performs this cleanup; the only interaction is the resurrection risk noted above.

Checklist

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