[Language] Expand conformance corpus to full construct coverage

[cssbruno]

Follow-up from #295 and PR #300.

Problem

#295 added the conformance corpus runner and a seed of seven accept/reject cases
(internal/lang/testdata/conformance/). The contract is now pinned where cases
exist, but coverage is partial: most blocks, g: directives, metadata keywords,
and endpoint forms have no corpus case, and nothing asserts that every documented
construct is exercised. docs/language/conformance.md describes the mechanism;
the breadth is still to fill.

Relevant code:

• internal/lang/conformance_test.go: the runner
• internal/lang/testdata/conformance/{accept,reject}
• docs/language/spec.md, docs/language/grammar.md: the constructs to cover
• lang.MetadataKeywords, view.SupportedDirectiveNames: construct registries

Scope

Expand the corpus to at least one accept case per stable/partial construct and
one reject case per planned/deprecated construct, and add a coverage assertion
that fails when a construct in the code registries has no corpus case. Optionally
add a machine-checked EBNF grammar and assert the corpus exercises every
production.

Acceptance Criteria

• Every metadata keyword, supported g: directive, supported top-level block,
  and endpoint form has at least one accept case.
• Every planned/deprecated construct has a reject case asserting its code.
• A coverage test fails when a registry construct has no corpus case.
• docs/language/conformance.md documents the coverage guarantee.

Verification

go test ./internal/lang

[cssbruno]

Expanded in #308, with a finding that bounds the scope: full per-construct accept coverage is infeasible through single-file CheckSource. Component state is a Go-typed contract (state SomeState NewSomeState), so reactive g: directives can't check clean without the Go types; act/api need exported handlers; layout/wasm/asset/css need sibling files or config.

The corpus now pins what one file can verify (package/metadata/route/view/style/literal build/slots + the rejection contracts), and docs/language/conformance.md documents that the rest need a project-level conformance harness. Suggest rescoping the remaining 'every construct has an accept case' goal to that separate project-level harness rather than the single-file corpus.

[cssbruno]

Completed in #309. Full construct coverage is now a verifiable gate rather than an aspiration: TestConformanceCoversEveryConstruct partitions every construct in the stability registry into corpusConstructs (single-file corpus) or integrationCoverage (a named integration test for project-context constructs that single-file CheckSource can't exercise). Adding a construct fails the test until it's classified, and a companion test rejects stale entries — so no construct can silently go untested. Will close on merge.