Shared test corpus

Golden input/expected pairs used by every language implementation. The corpus is the cross-language source of truth — if TS, Rust, and C disagree on a sample, the corpus says who is right.

Layout

Mirrors spec/:

corpus/
├── labels/10/POS/sample-001.json
├── labels/10/POS/sample-002-invalid.json
├── labels/44/POS/sample-001.json
├── wildcards/arinc_702/sample-001.json
└── ...

Format

{
  "spec": "labels/10/POS",
  "source": "acars-decoder-typescript/lib/plugins/Label_10_POS.test.ts",
  "description": "human-readable test case description",
  "input": {
    "label": "10",
    "text": "POS082150, N 3885,..."
  },
  "expected": {
    "decoded": true,
    "decoder": { "name": "label-10-pos", "type": "pattern-match", "decodeLevel": "partial" },
    "formatted": { "description": "Position Report", "items": [ ... ] },
    "remaining": { "text": "..." }
  }
}

Extracting from TS tests

acars-decoder-typescript's lib/plugins/*.test.ts files are the source of truth. A small extractor script (planned, codegen/scripts/extract-corpus.ts) instruments the Jest run to dump (input, decodeResult) pairs to JSON. Manually-curated samples are also welcome and should set source: "manual".

Validation

Each language repo's CI loads every JSON file, runs the decoder against input, and deep-equals against expected. Divergence fails CI in that repo.

A reference TypeScript interpreter (planned, codegen/src/interpret.ts) walks the spec IR directly and runs the corpus — this validates that specs are decodable before any language repo adopts them.