Add docTR as 5th model and tie-breaker in ROVER consensus

[Arkoniak]

Background

Manual audit of 2-vs-2 content disagreements (see docs/experiments/2026-06-29-surya-spell-checker-bias.md) showed that 4-model ROVER ties are currently broken by insertion order (Surya wins). This works well in practice (35/36 correct) but is brittle: all 4 VLMs carry language-model priors that push toward standard spellings, so on non-standard dialect or rare compound words a tie can go the wrong way.

The fix: add a 5th model that reads glyphs without language-model priors, turning 2-vs-2 into 3-vs-2.

Why docTR

• Two-stage CNN pipeline (detection + CRNN recognition) — no language model, no spell-checking bias
• Fast enough to run on the full book (1–2 s/page on GPU, also works on CPU)
• Gives word-level bboxes + confidence scores, compatible with our block structure
• Pure Python, pip-installable, no trust_remote_code

Plan

1. Write scripts/run_doctr.py following the same run layout as run_surya.py / run_unlimited.py — one manifest.json + per-page page_NNN.json with text, blocks (label/bbox/text per word), format: "doctr".

2. Run on full book with --label full (same suffix as the other models so consensus.py picks it up automatically).

3. Integrate into consensus.py as a tie-breaker: docTR participates in the full alignment (its tokens go into every column), but in vote_column() its vote only counts when the 4 VLMs are split 2-vs-2. Concretely: compute the 4-VLM majority first; if agreement == 0.5, use docTR to break the tie; otherwise ignore its vote. This preserves the 4-VLM consensus on unambiguous columns while giving an independent signal on contested ones.

4. Evaluate: re-run consensus, compare content_agreement_rate and check the former 2-vs-2 tie cases (especially Don'/Don't, creds/credits, raied/raised).

Acceptance criteria

• run_doctr.py produces a valid run directory consumable by consensus.py
• vote_column() accepts an optional tiebreaker model name and uses it only on 2-vs-2 splits
• Full-book consensus re-run with docTR as tie-breaker, results recorded in docs/experiments/
• content_agreement_rate does not regress; the known bad tie (p221 Don't) is resolved