[Env Request] LaTeX OCR — multimodal environment on the new Task API + dataset streaming

[adithya-s-k]

Summary

Add a multimodal (vision + text) dataset-backed environment: the agent is
shown an image of a mathematical expression and must return its LaTeX
source. Reward is computed server-side against the hidden ground-truth LaTeX.

This environment is built specifically on top of the newly introduced Task
API (#726) and adds dataset streaming for datasets too large to
materialize — the two capabilities this proposal is meant to exercise and
showcase end-to-end.

Why this env

1. Exercises the new Task API — a real, dataset-backed consumer of
   list_splits / num_tasks / get_task / get_task_range, so the Task API
   has a concrete reference implementation beyond the importer templates.
2. Dataset streaming — a first environment that streams a Hugging Face
   dataset instead of downloading it, so TB-scale datasets (millions of rows,
   hundreds of GB of images) can be served from a small Space.
3. Multimodal — image input + text (LaTeX) output, gradeable with any
   vision-LLM policy. Fills a gap: most reference envs are text-only.

Proposed design

• Episode (bandit): reset(split, index) → observation with base64 image
  (target hidden) → step(LatexOCRAction(latex=…)) → reward, done=True,
  ground truth revealed.
• Task API modes (LATEX_OCR_MODE):
  • materialize — split loaded + indexed; reset(split, index) random access.
    Best when the dataset fits on disk.
  • stream — sequential cursor over a streamed split, no full download;
    observations carry progress (index / total / remaining / pct_done);
    num_tasks comes from dataset metadata only; no-repeat within a pass.
    reset(index=i) random access is intentionally unsupported (streaming is
    sequential). This is the mode for very large / TB-scale datasets.
• Reward (LatexOCRRubric): 0.8·(1−CER) + 0.2·exact_match, whitespace-
  insensitive normalized character edit distance; pure-Python (no extra deps).
• Custom Gradio "Try it" tab: get a task image, type LaTeX, see the score.
• Any (image, latex) dataset works via LATEX_OCR_DATASET (+ column env vars);
  default is unsloth/LaTeX_OCR.

Scope / non-goals

• No new core APIs — uses the existing Environment + the new Task API as-is.
• Semantic/rendered reward (render-and-compare) is out of scope for v1.

Validation

Verified end-to-end locally and on a deployed HF Space: Task API, streaming
cursor + live progress, no-repeat, and a real vision-LLM policy via the HF
Inference Router (mean reward ~0.77, incl. exact matches). Rubric has unit tests.