feat(parser): re-architect ReceiptParser into a rule/engine pipeline

Re-architect `ReceiptParser` from one ~320-line static class into the rule/engine pipeline sketched in the receipt-parser design notes, scoped to restaurant/bar/cafe receipts. ## Design - Decision: `docs/adr/0006-receipt-parser-pipeline.md` - Detector/rule/score reference: `docs/15-receipt-parsing.md` Every sub-issue inherits these plus the invariants below; they are the durable home for the detail (regex catalogue, layouts, scoring, OCR-misread maps), so the issues stay terse and point at the docs. ## Goal - Cleaner, extensible parser: small named rules instead of one growing method. - Parse more real receipts correctly, each new behaviour proven by a fixture. ## Target shape `OcrResult` -> normalize -> classify lines -> item rules + engine -> bill detectors + engine -> validators -> `ParsedReceipt`. New code under `BillSplitter.Domain/Parsing/**`, pure, no new dependencies. ## Invariants (every sub-issue) - The 33-fixture corpus stays green. An existing `expected.json` changes only when a capability deliberately improves that receipt, with justification. - Facade `ReceiptParser.Parse(OcrResult)` stays static -> `OcrWorker.cs:70` never changes. - Money stays `long` minor units; the server owns all math. - No new persistence; discount/negative lines stay parked in `Warnings`. - Diagnostics are in-memory / test-only. No receipt text in prod logs. - Contract docs updated in the same PR (docs/06 parsing, docs/11 testing). No wire-contract change is expected, so docs/03-05 stay untouched. ## Phase A - structural extraction (behaviour-preserving, one PR per layer) - [x] A1 scaffold parsing pipeline skeleton and static facade - [x] A2 extract text normalizer - [x] A3 extract line classifier - [x] A4 extract item rules and selection engine - [x] A5 extract bill detectors and engine - [x] A6 add in-memory parse-decision trace ## Phase B - capabilities (each ships with new fixtures) - [x] B7 warn when receipt totals do not reconcile - [x] B8 fix OCR character misreads in money spans - [x] B9 attach modifier lines to their item - [x] B10 merge wrapped item names across lines (gated) - [x] B11 reorder scrambled lines by bounding box (last, riskiest) - [x] B12 de-duplicate repeated receipt copies Sequencing: A1->A5 in order, A6 after A5; then B7/B8/B9, then B10, then B11, then B12. ## Out of scope - Weight-priced, fuel, and supermarket layouts. - Multi-currency / locale-decimal expansion beyond the current corpus. - Modelling discounts in the domain - its own future epic (`Bill.DiscountMinor` plus Redis, DTO, docs/02-06). - Any prod logging of parse decisions.

TimCane commented

2026-07-07 09:24:15 +01:00

(Migrated from github.com)

Re-architect ReceiptParser from one ~320-line static class into the
rule/engine pipeline sketched in the receipt-parser design notes, scoped to
restaurant/bar/cafe receipts.

Design

Decision: docs/adr/0006-receipt-parser-pipeline.md
Detector/rule/score reference: docs/15-receipt-parsing.md

Every sub-issue inherits these plus the invariants below; they are the durable
home for the detail (regex catalogue, layouts, scoring, OCR-misread maps), so
the issues stay terse and point at the docs.

Goal

Cleaner, extensible parser: small named rules instead of one growing method.
Parse more real receipts correctly, each new behaviour proven by a fixture.

Target shape

OcrResult -> normalize -> classify lines -> item rules + engine -> bill
detectors + engine -> validators -> ParsedReceipt. New code under
BillSplitter.Domain/Parsing/**, pure, no new dependencies.

Invariants (every sub-issue)

The 33-fixture corpus stays green. An existing expected.json changes only
when a capability deliberately improves that receipt, with justification.
Facade ReceiptParser.Parse(OcrResult) stays static -> OcrWorker.cs:70
never changes.
Money stays long minor units; the server owns all math.
No new persistence; discount/negative lines stay parked in Warnings.
Diagnostics are in-memory / test-only. No receipt text in prod logs.
Contract docs updated in the same PR (docs/06 parsing, docs/11 testing). No
wire-contract change is expected, so docs/03-05 stay untouched.

Phase A - structural extraction (behaviour-preserving, one PR per layer)

A1 scaffold parsing pipeline skeleton and static facade
A2 extract text normalizer
A3 extract line classifier
A4 extract item rules and selection engine
A5 extract bill detectors and engine
A6 add in-memory parse-decision trace

Phase B - capabilities (each ships with new fixtures)

B7 warn when receipt totals do not reconcile
B8 fix OCR character misreads in money spans
B9 attach modifier lines to their item
B10 merge wrapped item names across lines (gated)
B11 reorder scrambled lines by bounding box (last, riskiest)
B12 de-duplicate repeated receipt copies

Sequencing: A1->A5 in order, A6 after A5; then B7/B8/B9, then B10, then B11,
then B12.

Out of scope

Weight-priced, fuel, and supermarket layouts.
Multi-currency / locale-decimal expansion beyond the current corpus.
Modelling discounts in the domain - its own future epic (Bill.DiscountMinor
plus Redis, DTO, docs/02-06).
Any prod logging of parse decisions.

Rows
Columns

feat(parser): re-architect ReceiptParser into a rule/engine pipeline #77