[Feature]: one test that pins graph.json invariance for every extractor (checkout folder, rebuild, line endings) #4217

Problem or use case

The same source tree should always give the same graph.json. Seven recent bugs broke that in one of four ways, and each was found by diffing two builds by hand:

Each fix added a narrow test for its own extractor or write path. Nothing checks the invariant across all extractors, so the next extractor, or the next change to a write path, can bring the same bug back without any test failing.

Proposed solution

One test file that builds the whole tests/fixtures corpus (every language with an extractor) through the real CLI, in --no-cluster and clustered mode, and compares every node and edge with all attributes:

• two copies in different folders give the same graph;
• the checkout folder name appears nowhere in graph.json;
• a cold build and a rebuild of the unchanged tree are the same (top-level keys too);
• an LF copy and a CRLF copy give the same graph;
• every raw edge endpoint is a declared node;
• the corpus is not trivially empty (guards against a silent empty build).

Community ids are left out of the comparison (they come from clustering; #3761 covers that).

Alternatives considered

• Keep adding one narrow test per bug: that is what happened seven times, and it only protects the one extractor each bug was found in.
• A CI job that builds a real repo and diffs: slower, needs network, and harder to run locally than a pytest file on the fixtures already in the repo.

Area

Developer experience or CI

Are you willing to submit a PR?

Yes, I can work on this

Compatibility / behavior impact

Test only. No change to node/edge ids, cache, graph.json schema, skills, MCP output or incremental updates. Adds about 20 s to the suite.

Additional context

Written and checked against 0.9.80 (6478eb7): 12/12 pass. Against 0.9.79 the three checks covered by #4180/#4182/#4188 fail, which is the point. PR right after this issue. Windows 11, Python 3.12.12.