Add token export mask visualizer

[samsja]

Summary

• Add scripts/token_export_mask_visualizer.py for rendering static HTML from token export JSONL files.
• Compare DPPO prob_delta masking against IcePop importance-ratio masking.
• Include token-level KL gradient, DPPO/IcePop overlay toggles, live threshold tuning, quantitative mask/disagreement counts, and top-token examples.
• Support file, step-directory, and token_exports-root inputs with optional rank/env filters and optional tokenizer decoding.
• Add docs/token-export-mask-visualizer.md with quick-start commands, common options, mask-rule notes, UI color meanings, and an agent workflow.

Validation

• python -m py_compile scripts/token_export_mask_visualizer.py
• uvx ruff check --config=pyproject.toml
• uvx ruff format --check --config=pyproject.toml
• Smoke render on /beegfs/outputs/glm5.1-rlm-prod/token_exports --step 8 --rank 0 --top-records 1 --window-tokens 20 --no-decode

---

Note

Low Risk
Low risk: changes are additive (new debugging/visualization script + docs) and only tweak formatting config, with no impact on training/runtime code paths.

Overview
Adds scripts/token_export_mask_visualizer.py, a CLI tool that scans token export JSONL (file/step_N dir/token_exports root), computes DPPO vs IcePop mask decisions, and generates a self-contained HTML page with KL heatmap, mask overlays, record filtering, live threshold tuning, and top-token example tables.

Documents the tool in docs/token-export-mask-visualizer.md and links it from the docs index/navigation. Updates pyproject.toml to exclude the new script from ruff format.

^{Reviewed by Cursor Bugbot for commit 038cdaf. Bugbot is set up for automated code reviews on this repo. Configure here.}