0.3.0 — structural tables

Structural tables: documents are no longer flattened before the model sees them.
Everything here was tuned against a 314-file corpus of real invoices and then
measured for accuracy against 76 Factur-X documents with embedded EN16931 ground
truth, rather than against intuition. Two of those measurements changed the design
and one stopped a change from shipping — details inline.

Fixed

• Trailing-minus accounting notation in LenientDecoding.parseDecimal. SAP /
  German invoice amounts print the sign after the digits ("1,12 -", "12-"), and
  a real corpus document does exactly that. Those forms now decode as negatives
  (−1.12, −12) instead of being rejected. Optional whitespace before the sign is
  allowed; a trailing + is accepted as an explicit positive. Two-sign forms
  (-12-, 12--, +12-, parenthesised-plus-trailing) stay rejected, as do the
  existing multi-dot / multi-sign / junk-exponent hardenings, pinned as a table so
  the boundary is visible in one place. Parenthesised negatives and the scientific
  path are unchanged.

Added

• Structural tables surface to the model and callers (stage 2). Detected tables
  from geometric reconstruction are returned on ExtractionResult.tables and may be
  appended to the extraction prompt as a labelled Markdown section (linear document
  text is left unchanged — additive, not a substitute). When no tables are found or
  tableDetection is .off, the prompt is byte-identical to the pre-feature shape.
  Chunked runs assign whole tables by page (never a half table); if page filtering
  would drop every table, the full set is attached to the first chunk. Docs:
  API, Examples, README limitations.

Changed

• Schema-gated table prompt injection under .automatic. Always appending
  detected tables to the prompt hurt header-field accuracy on types without
  collections (measured −11.4 pp header accuracy on 76 invoices) while helping
  line-item structure when the schema has arrays. .automatic still runs geometric
  detection and always exposes grids on result.tables; the prompt section is
  included only when the target extractionSchema contains a collection (array)
  anywhere, including nested. Header-only types get a prompt byte-identical to
  .off. No new enum case; callers who need grids without line items still use
  result.tables.

  Measured on 76 Factur-X invoices with embedded EN16931 ground truth, same local
  MLX model per run, temperature = 0, tables .automatic vs .off:

  	Qwen2.5 1.5B 4-bit	Qwen2.5 7B 4-bit
  header accuracy, tables off	62.7%	95.2%
  header accuracy, tables on	51.3%	95.8%
  Δ on the files where tables were injected	−11.4 pp	+0.8 pp
  Δ line description	−1.0 pp	+7.1 pp

  So the distraction is a small-model artefact, not a property of the approach:
  at 7B the table section no longer costs header accuracy and clearly helps line
  descriptions. Schema gating is kept because it is exactly right for weak models
  and for header-only schemas, and costs a strong model nothing. On header-only
  types the two arms produced byte-identical extractions on all 76 documents —
  only the reported tables count differs, which is the intended behaviour.

• Table detection precision against real invoices. TableDetector now splits
  multi-column regions on large vertical gaps (line items vs totals), bridges short
  single-column description lines under items, keeps only “spine” rows (numeric /
  header-like) when clustering columns, drops sparse and all-empty columns, and
  requires fill density ≥ 0.55 before emission. On a 314-file invoice corpus this
  raised median density from ~0.71 to 1.0, cut pdf/ average tables/file from 3.4
  to 1.8, and eliminated all-empty columns, while preserving line-item recall on
  Coolblue, Sammy Maystone, fixtures/invoice.pdf, and fixtures/receipt.png.
  Unit tests cover a Coolblue-style merged-region layout and an empty-column case.

• Recursive XY-cut (horizontal column bands). Before row grouping, the detector
  splits a page on a vertical whitespace corridor — the largest mid-X gap between
  blocks — when the gap is ≥ 0.06, ≥ 1.25× each band’s internal mid-X structure,
  both bands have ≥ 2 multi-column rows, and their Y-ranges overlap (≥ 25% of the
  shorter band). Each band then runs the existing vertical region split. This
  recovers side-by-side documents that previously merged into one sparse mega-grid
  (or density-rejected to zero), e.g. hard/invoice_table_detect_img1.jpg, without
  bisecting single-document line-item tables (Coolblue, Sammy). Unit test covers two
  independent 3-column grids separated by a wide corridor.

• XY-cut corridor vs table-gutter discrimination. Width alone cannot tell a
  vertical region boundary from a table’s inter-column gutter, so pure XY-cut
  shattered genuine line-item grids (e.g. Kostenrechnung 4×6 → six half-width
  fragments; Hetzner 8×4 → 8×2/4×2/11×2). A candidate corridor is now kept
  only when multi-column row baselines across the gap largely fail to align: the cut
  is refused when the shorter multi-col side’s mid-Y match is ≥ 0.80 (strong
  co-tabular signal, even if one side has extra chrome rows) or when both directed
  matches are ≥ 0.50 (balanced shared grid). Side-by-side documents with unrelated
  line positions still split (hard/invoice_table_detect_img1.jpg → 4×4 + 6×5).
  On the 90-file pdf/ set, line-item-shaped grids (rows ≥ 3, cols 3–6, density
  ≥ 0.80) recover to 36/90 (40%) from 32/90 under pure XY-cut, matching the
  pre-horizontal-cut tightening rate, while median density stays 1.0, p90 columns
  5, and pdf/ avg tables/file 1.79. Unit test pins a dense 6-column grid staying
  one table against two misaligned 3-column grids staying two.