Cross-grammar sync tests for the refs grammars

[gandarfh]

What

Shared-corpus sync tests asserting the Lezer and tree-sitter refs grammars agree on the same inputs, plus a CI job that runs them on every PR.

Design

• 11 corpus cases covering aliases, dot paths, array indexing, env vars, surrounding text, and two error-recovery cases marked .diverges. where engines may legitimately differ.
• Both engines are normalized to one canonical node vocabulary (normalize.ts); the tree-sitter side shells out to the CLI so no native bindings or WASM build are needed.
• The assertion encodes the sync contract: the Lezer stream must be a coarsening of the tree-sitter stream (every Lezer node exists there with the same span and kind — Lezer may emit less structure, never different structure), and ref anchors must match exactly in both directions so a degenerate empty Lezer parse cannot pass.
• Contiguous text nodes are coalesced before comparison: the engines chunk plain-text runs differently and the segmentation carries no rendering meaning.

Fixes found while wiring it up

• tree-sitter parse --quiet suppresses the tree output on CLI 0.26, so the tree-sitter side returned nothing; the flag is gone.
• The CLI exits non-zero when the tree contains ERROR/MISSING nodes (exactly the error-recovery corpus); the runner now keeps the tree printed to stdout instead of discarding it.

Validation

11/11 locally against tree-sitter CLI 0.26.8; the new cross-grammar CI job builds the Lezer package and runs the suite on ubuntu.