Commercial real estate invoice intake and reconciliation

[maccora]

Summary

Adds a self-contained subsystem under docex/intake/ that reads commercial
real estate invoice PDFs of arbitrary layout, extracts their fields, and
reconciles them against recorded ground-truth lease actuals (base rent, CAM,
real estate tax and insurance recoveries, totals, square footage, pro-rata
share) so an overcharge can be caught before it is paid.

Design

• Cost-ordered extraction cascade. A free, deterministic heuristic runs
  first; an optional caller-provided LLM is invoked only for unresolved or
  disputed fields. A clean invoice costs zero LLM calls. The LLM is a BYO
  callable, so the core takes no hard provider dependency.
• Self-improving learning loop. Label phrasings confirmed against ground
  truth are recorded and promoted into the heuristic's alias set, so the cheap
  tier absorbs vendor formatting over time and the LLM is needed less.
• Reconciliation. Type-aware tolerances (a cent on money, configurable on
  dates) with per-charge comparison by category. Verdicts are matched,
  discrepancy, incomplete, or unresolved.
• Ground-truth retrieval. Records are matched by stable identifier
  (invoice number, then PO). An optional embedding-similarity matcher provides
  fuzzy retrieval to find the closest lease when an invoice has no clean
  identifier; it never overrides reconciliation, so a wrong match shows up as a
  discrepancy rather than being silently trusted.
• DocEX integration. An InvoiceIntakeProcessor wraps the pipeline as a
  BaseProcessor; the core pipeline itself carries no database dependency.

A note on a deliberate decision: an embedding-similarity tier was considered for
field extraction and left out, because the learning loop already removes its
only benefit and it carries a false-positive risk the cascade cannot cheaply
audit. The rationale is documented in docex/intake/extractors/cascade.py and
the package README.

Testing

• Unit coverage of every layer (normalization, charge taxonomy, heuristic
  extraction traps, reconciliation tolerances, matching, learning).
• Randomized invoice scenarios across varied layouts, labels, and values.
• The self-improving loop verified end to end.
• LLM tier covered by deterministic stubs plus a live test gated on
  ANTHROPIC_API_KEY.
• Two realistic invoice PDFs committed under example_docs/cre_invoices/ (a
  positive one that reconciles, and a negative one overstating CAM that is
  flagged), run through the full pipeline including pdfminer.

165 tests pass; lint is clean. New optional dependencies are confined to the
existing pdf extra (pdfminer) and dev/test tooling (reportlab).

See docex/intake/README.md for the full design, testing notes, and the
assumptions the intake makes about its inputs.