ocr: tidy CLAUDE.md (info dump → conventions + status + gotchas; history → OCR-BENCHMARK.md) (#68)

* ocr: tidy CLAUDE.md — hoist conventions, split history out (949->238 lines)

CLAUDE.md had grown to ~950 lines / 60KB — an info dump interleaving durable conventions,
per-script gotchas, and a benchmark changelog, so the reference material was hard to find.

- New "Conventions & invariants" block at the top (single-file/pinned-dep, vLLM-image +
  preflight, env guards, context-length invariant, collision guard, image bounding, dep
  sanity) — the single source the planned self-review skill enforces.
- Compact "Script status" table + tightened per-script gotchas, keeping every load-bearing
  quirk (surya site-packages path, nanonets :v0.10.2, dots content_format + bbox mapping,
  unlimited dedicated image, lift naming, deepseek crop-dim bug, the pyarrow->datasets lesson).
- Benchmark result tables (Run 1/2/3, cross-dataset, judge/human-eval history) moved to a new
  OCR-BENCHMARK.md; CLAUDE.md keeps the tooling how-to + a headline.
- Sparks -> issue links (#67 bucketbag; self-review skill; incremental-uploads superseded by
  Buckets/bucketbag); deferred follow-ups consolidated (ALTO export, multi-page SGLang recipe).

No load-bearing gotcha dropped (verified by keyword sweep); no private names.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ocr: note that model pins are temporary and should be loosened over time

Add a Conventions bullet: image/version pins (:v0.10.2, :v0.20.1, nightly,
surya-ocr==0.20.0) are workarounds for current ecosystem gaps; record why each
exists + what would loosen it, and re-test/relax over time (that's what the
bump-vllm-pins skill is for).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ocr: restore 3 dropped gotchas + clarify deepseek bug provenance (review follow-up)

Adversarial drop-check on the tidy found three 'don't re-debug this' notes
that were lost in the condensation; restored: nightly-index transient
failures (ARM-only wheels -> retry), glm blank-page chattiness is model
quality not a crash, and the dots content_format fix not applying to v1.
Also reworded the deepseek crop-dim bug to say it was hit on nightly
(the script now resolves stable wheels, per the status table).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: davanstrien