dev-env: a stale venv can silently test a different tree-sitter grammar than CI, inverting parser results #1625

Hit this while reviewing #1610 and it produced a confidently wrong review, so it's worth a guard.

pyproject.toml pins:

tree-sitter-language-pack>=1.6,<2.0

A venv created before that pin was raised can still hold 0.13.0 and nothing complains — pytest runs, the suite passes, and parser behaviour differs materially from CI.

Why it matters

The Kotlin single-line-object misparse (#1600) is entirely grammar-dependent:

grammar 1.14.3 (pinned):  object A { fun x() = 1 }  ->  infix_expression, 0 object_declaration captures
grammar 0.13.0 (stale) :  object A { fun x() = 1 }  ->  object_declaration, captures = ['A']

So on a stale venv the bug does not reproduce, and #1610 looks like a no-op that can be closed. I nearly did exactly that. The reporter and I would have disagreed indefinitely with both of us running pytest and both seeing green.

This cuts the other way too: a parser fix developed against 0.13.0 could pass locally and fail in CI.

Suggested guards

Any one of these would have caught it:

1. A session-scoped test assertion that the installed tree_sitter_language_pack.__version__ satisfies the pin, failing loudly rather than skipping. Cheap and self-maintaining.
2. tests/run_tests.sh verifying installed versions against pyproject.toml before running (e.g. pip check plus an explicit version compare) — pip check alone does not catch this, since the stale package satisfies no broken dependency graph, it just predates the pin.
3. Documenting a pip install -e . --upgrade step in the contributing guide for anyone with an older checkout.

(1) is the most robust because it fails on the machine that's actually wrong.

Related: my local venv also had codegraphcontext 0.4.16 installed while the repo was at 0.5.7, which made cgc --version report 0.4.16 — a false positive for #1262. Same root cause: a long-lived venv drifting from the repo.