Prune eval/ to the recreation core

[chonknick]

Trims eval/ from 81 → 50 files, keeping what's needed to recreate/reproduce, dropping one-offs.

Kept:

• Fixture set: html/ (13), reference/ (12), output/ (13), results/ (7 per-method scores), manifest.json, README.md
• 3 canonical harnesses: eval_latte_large_vs_dripper.py (ROUGE-5+qrater, per the GPU runbook), bench_all_methods.py, eval_all_methods_200.py

Removed:

• ~24 one-off diagnostic/inspect/compare/bench-variant scripts (exploratory probes; narrative in work.log)
• 7 loose regenerable artifacts (*_results.json, a .log, qrater_scores.json)

Kept scripts compile, are self-contained (no refs to removed files), and import the chunker from the pulpie package. eval/README.md now documents the harnesses.

[Copilot]

Pull request overview

This PR prunes the eval/ directory down to a smaller “recreation core” by removing one-off/diagnostic evaluation scripts and regenerable artifacts, while keeping the canonical harnesses, fixtures, and benchmark score outputs needed to reproduce reported numbers.

Changes:

• Removed numerous exploratory/diagnostic eval scripts and generated artifacts (*_results.json, logs, old qrater score dump).
• Kept the evaluation fixture set (html/, reference/, output/, results/, manifest.json) and the canonical harness scripts (per PR description).
• Updated eval/README.md to document the remaining directories and harness scripts (and to point to tests/gpu/AGENTS.md for the canonical runbook context).

Reviewed changes

Copilot reviewed 30 out of 32 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
eval/README.md	Documents the retained eval directories and the canonical harness scripts used to reproduce benchmark numbers.
eval/test_strip_tail.py	Removed a one-off experiment script (strip <marked-tail> annotations and compare ROUGE).
eval/test_empty_pages.py	Removed an exploratory script analyzing LLM “empty page” labeling and potential training implications.
eval/test_deepseek_pipeline.py	Removed a DeepSeek+MinerU-HTML pipeline validation script for WebMainBench.
eval/score_qrater.py	Removed a local-path qrater scoring script (model path tied to a specific workstation).
eval/qrater_scores.json	Removed a generated qrater score output artifact.
eval/inspect_eval_pages.py	Removed an inspection tool for comparing pulpie output vs ground truth per page.
eval/inspect_empty_pages.py	Removed an investigation script for empty-output pages, including feature export probing.
eval/eval_sample50.py	Removed a 50-page sampling harness that extracted and qrater-scored outputs.
eval/eval_sample50_results.json	Removed a generated 50-sample results artifact.
eval/eval_qrater_dripper.py	Removed a multi-method qrater clean-rate comparison harness (pulpie/dripper/orange/raw).
eval/eval_qrater_clean_rate.py	Removed a qrater clean-rate evaluation script (wmb/cc sources).
eval/eval_dripper_native.py	Removed a Dripper native-vLLM evaluation script and comparison report.
eval/eval_deepseek_wmb.py	Removed a DeepSeek-on-WMB evaluation script with concurrency and scoring.
eval/eval_block_classifier.py	Removed a WebMainBench eval script for the [BLOCK] token classifier approach.
eval/error_analysis.py	Removed an error analysis script categorizing failure modes and output distributions.
eval/diagnose_rouge.py	Removed a diagnostic script exploring causes of low ROUGE-5 (empties, formatting, etc.).
eval/diagnose_rouge_gap.py	Removed an in-depth diagnostic script analyzing the ROUGE gap vs Dripper (oracle/all-main baselines, divergences).
eval/compare_formatting.py	Removed a one-off ROUGE comparison script with/without markdown formatting.
eval/block_classifier_wmb_results.log	Removed a logged output artifact from a prior block-classifier WMB run.
eval/bench_throughput.py	Removed a throughput benchmark script for the [BLOCK] classifier with batching.
eval/bench_rouge.py	Removed a ROUGE-5 benchmark script against WebMainBench using a fast custom scorer.
eval/bench_rouge_results.json	Removed a generated ROUGE benchmark results artifact.
eval/bench_rouge_h2t.py	Removed a ROUGE benchmark variant using --html + html2text canonicalization.
eval/bench_rouge_h2t_results.json	Removed a generated results artifact for the html2text-canonicalized benchmark.
eval/bench_qwen_classify.py	Removed an async vLLM-based Qwen classifier benchmark script (Dripper-style pipeline).
eval/bench_qwen_classify_results.json	Removed a generated results artifact for the Qwen benchmark.
eval/bench_pulpie_gpu_only.py	Removed a GPU-only inference throughput benchmark script.
eval/bench_pulpie_gpu_compile.py	Removed a GPU throughput benchmark variant using torch.compile.
eval/bench_pulpie_a100.py	Removed an A100 end-to-end throughput benchmark script (Extractor vs Pipeline).
eval/analyze_failures.py	Removed a small script for inspecting dirty pages from the 50-sample eval.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.