[faresmohamed260]

V3 nominal boundary-correction follow-up

Records the held-out correction-pattern candidate sweep and the corrected-candidate classifier gate that followed the merged nominal qualification record.

New evidence

• train-book-only boundary corrections generalize to unseen books;
• top-4 corrections: 190,308 candidates / 72.84% exact-span ceiling;
• top-8 corrections: 318,366 candidates / 78.05% exact-span ceiling;
• top-8 slightly beats the earlier brute ±3/max-10 ceiling (77.84%) at roughly half the candidate volume;
• however, the lightweight corrected-span rankers do not beat the simpler exact-noun-chunk classifier:
  • exact noun chunks: OOF AP 0.8115 / best F1 0.6039
  • top-4 corrected: OOF AP 0.6482 / best F1 0.5575
  • top-8 corrected: OOF AP 0.5322 / best F1 0.5060

Decision

• keep exact noun chunks + lightweight held-out classification as the leading simple public nominal baseline;
• reject the tested top-4/top-8 corrected linear classifiers;
• retain learned correction patterns only as measured candidate-generation evidence for a stronger direct span model;
• do not introduce any production nominal/identity threshold from these scores.

Next challenger

A direct span-structure learner (preferably spaCy SpanCategorizer or equivalent compact local model) should be benchmarked before escalating to a larger encoder.

Documentation-only PR. No runtime dependency, production provider, or canon policy changes.