[Language] Add a machine-checked .gwdk conformance corpus

[cssbruno]

Problem

The .gwdk language contract is documented as prose in docs/language/grammar.md
and docs/language/spec.md, but nothing executable pins it. There is no
conformance corpus of accept/reject samples with golden diagnostics, and no
machine-checked grammar verified against the parser. The contract can therefore
drift silently: a parser change can accept or reject syntax that the spec still
describes the old way, and no test catches it.

For a milestone literally named "Compiler + Language Contract," a frozen,
CI-run conformance corpus is the artifact that makes the contract real.

Relevant code:

• docs/language/grammar.md, docs/language/spec.md: prose contract
• internal/parser, internal/lang: parse + check entry points
• testfixture/, internal/*/testdata: scattered existing fixtures

Scope

Introduce a versioned language conformance corpus:

• accept/ samples that must parse and validate clean.
• reject/ samples that must produce a specific stable diagnostic code (and,
  where available, span).
• Golden diagnostics regenerated deterministically by a test runner.
• A doc mapping each grammar production / spec section to at least one fixture.

Optionally add a machine-checked grammar (EBNF) and assert the corpus exercises
every production. Keep the corpus the cited source of truth in the spec.

Acceptance Criteria

• A conformance corpus directory exists with accept and reject cases.
• Each reject case asserts a stable diagnostic code; golden output is
  deterministic and regenerable.
• A CI test runs the corpus and fails on drift.
• docs/language/spec.md and grammar.md reference the corpus as the contract
  source of truth.
• Adding new syntax requires a corpus entry (documented in
  docs/compiler/syntax-contributors.md).

Verification

go test ./internal/lang ./internal/parser
# plus the new corpus runner package

[cssbruno]

Implemented and merged in #300.