Engine-agnostic morphological-parser conformance suite (v2: typologically-selected languages)

[johnml1135]

Engine-agnostic morphological-parser conformance suite (v2: typologically-selected languages)

An engine-agnostic conformance test suite for what a morphological/phonological parser must get right — a portable "all-language parser oracle". The C# HermitCrab engine is the founding oracle that generated ground truth; any engine implementing conformance/PROTOCOL.md's adapter contract can be validated against it.

This PR began as the F0–F3 fixture-per-mechanism suite and has been restructured into v2: a small set of dense, typologically-selected synthetic languages, replacing the 41 hand-minimized single-mechanism fixtures. Full design + rationale: docs/conformance-language-suite-plan.md.

Why restructure

The F0–F3 suite worked but (1) sprawled to 252 files (up to 7 per fixture, each word's story smeared across four), (2) made isolation the default when C#'s real bugs cluster in interactions, and (3) let coverage follow engine class-names rather than language phenomena — six constructs sat at zero coverage because no bug happened to hit them.

The v2 shape

• 8 synthetic languages (conformance/languages/) + 8 edge-cases (conformance/edge-cases/), exactly two files each (grammar.xml + words.yaml). 252 → 40 files.
• words.yaml is the single ground truth: per-word note + parses (signature/gloss/rules/exercises) or expect_fail/expect_skip, plus provenance, guess, budget_ms/expect_crash.
• Languages simulate typological families — Kirenti (Turkic), Zemreth (Semitic), Velamba (Bantu), Tavolan (Austronesian), Nuvattak (Inuit), Chaqru (Quechua), Ketselan (Athabaskan), Vetruna (Latin) — each dense enough that ordinary words exercise several constructs at once.
• All grammars are fully synthetic — invented lexemes, invented (plausible) segment inventories, no data from or orthographic claim about any real language. inspired_by/sources name only the phenomenon class each mechanism simulates (e.g. "Swahili/Sena-style reduplication"); no real-language grammar (Sena, Indonesian, Amharic, or any other) is included in this suite.

Coverage: nothing regressed, five zero-coverage constructs closed

• 19/19 in-scope constructs covered (Tracing is out of scope by design — never in expected.tsv's domain). The five previously-zero constructs now have homes: reduplication (ReduplicationHint), ModifyFromInput/InsertSimpleContext, CopyFromInput/InsertSegments, stratum ordering, and Guesser/LexicalGuess.
• The rules: field is a verified coverage dimension: self-check runs the oracle with tracing on and FAILs if the traced rule applications diverge from the declared list.
• rules.csv: zero dead rules in every grammar, with one principled exception — prule4 in the expect_crash cascade edge-case, which is un-exercisable by design (the engine crashes with InfiniteLoopException before any parse completes, so no word can ever attribute it).

The migration was mechanically proven before anything was deleted

conformance/parity-check.py (committed, re-runnable, exits nonzero on any violation) enforces the three migration floors + an absolute construct check, and self-validates that it discovered exactly 41 v1 fixtures. It gated the deletion of the v1 tree:

[PASS] A. provenance: 41/41 v1 strings carried verbatim
[PASS] B. per-construct floor: 14/14 v1 constructs covered in v2
[PASS] C. XAmple floor: 11/11 requires:[] constructs have a requires:[] v2 home
[PASS] D. absolute coverage: 19/19 in-scope constructs covered

It caught a real gap during authoring (the G1 pilot never carried three cooccurrence/* provenance strings) which was fixed before green. docs/conformance-migration-ledger.md is retained permanently as the v1→v2 provenance bridge; every v1 manifest.json.provenance appears verbatim in some v2 provenance: field.

Both run modes verified

• self-check (in-process C# oracle): 16/16 v2 fixtures PASS.
• adapter mode (materialize words.txt/expected.tsv → subprocess → diff, via hc-dotnet-wrapper.sh): 16/16 PASS. Running it surfaced (and this PR fixes) a materializer gap — an InvalidShapeException-skipped word was materialized as status ok instead of SKIPPED; the new expect_skip key pins that status and both modes now verify it consistently.
• HermitCrab unit tests: 67/67. csharpier clean.

Harness additions

verified traced-rules: self-check; --coverage-report now reports absolute construct coverage against constructs.txt (v2-native, survives the v1 deletion); --propose (prints a words.yaml patch on signature mismatch, never writes); expect_skip status; conformance/parity-check.py. The adapter wire contract in PROTOCOL.md is unchanged.

Docs

One conformance/README.md (≤120 lines); PROTOCOL.md trimmed (contract untouched); the v1 plan, HISTORY-MATRIX, and implementation notes archived under docs/archive/.

🤖 Generated with Claude Code

---

This change is 

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 73.34%. Comparing base (d9deb16) to head (4c79ed0).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #454      +/-   ##
==========================================
+ Coverage   73.33%   73.34%   +0.01%     
==========================================
  Files         443      443              
  Lines       37214    37214              
  Branches     5110     5110              
==========================================
+ Hits        27290    27294       +4     
+ Misses       8801     8798       -3     
+ Partials     1123     1122       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.