fix(graph): populate bundle metadata languages (#1516) - #1613

Fixes #1516. Every exported bundle — including the published registry bundles (#1180) — advertised no languages. There turned out to be two independent defects; the issue describes the first.

1. Nothing ever set f.language

cgc_bundle.py derives the list from a language property on File nodes, but no writer set it and the Kùzu File table had no such column. The if record["language"] guard then filtered every row out.

The per-file language is already known at write time as file_data['lang'], so this persists it. Three places were needed, not one:

• writer.py — set f.language on the File MERGE
• database_embedded_kuzu.py:173 — add the language STRING column
• database_embedded_kuzu.py:327 — a migration entry, so existing local databases gain the column rather than silently lacking it
• database_embedded_kuzu.py:821 — add language to SCHEMA_MAP['File']

That last one matters: Kùzu filters node properties against that allow-list and drops unknown ones silently, so the schema column alone would not have been enough and the failure would have been invisible. One of the tests targets exactly that step.

2. Unscoped exports skipped the block entirely

Even with the property populated, bundle export foo.cgc (no --repo) still produced no languages — the whole block sat inside if repo_path and repo_path.exists():, so the key was never set at all. That's why the metadata showed None rather than [].

Moved it out of that branch with an unscoped query variant.

Verification

Real bundles, KuzuDB backend:

before:  languages: None                        (Kotlin repo, both scoped and unscoped)

after:   bundle export lang.cgc                 -> ['kotlin']
         bundle export lang2.cgc --repo ./kt    -> ['kotlin']
         README.md                              -> "- **Languages**: kotlin"

mixed repo (a.py, b.kt, c.js):
         languages -> ['javascript', 'kotlin', 'python']

Tests

Three regression tests through the real GraphWriter path, including one specifically covering the allow-list drop. Verified all three fail on main and pass with the fix.

tests/unit/         1181 passed, 7 skipped
tests/integration/    44 passed