report: shorten production-path section; add industry-context subsection

- Halve the "Observations on a candidate production path" section.
  Collapse the 5 caveats into 4 (turnkey-maintenance point folded into
  the hallucination caveat), tighten the deployment-mode bullet to a
  pointer at the residency section, and replace the 3 next-step bullets
  with one closing line.
- Add an "Industry context — this pattern is 'document anchoring'"
  subsection inside the E08 deep dive (after Net interpretation).
  Names olmOCR (arXiv:2502.18443, Feb 2025) as the academic reference
  for the OCR-markdown + page-image + VLM pattern; lists commercial
  vendors running variants (Reducto, LandingAI ADE, LlamaParse Premium,
  Tensorlake, Google Document AI Layout Parser v1.6). Flags the
  olmOCR-2 / RolmOCR direction-of-travel away from anchoring with
  stronger VLMs, but notes the OCR leg is still load-bearing here
  (E08 vs E04 +8.6 pp F1.mean). Links to the new research doc.
- Track experiments/hybrid-model-research.md (referenced from the new
  subsection; companion to the existing content-understanding-research.md).

Author: strukalex