test(java): suite is order-dependent — 21/21 in isolation, 15 failures in full-suite runs (global JAVA_HOME mutation / shared cache suspects) #258

PROBLEM

The Java test family is order-dependent: results vary wildly between invocations of the same code.

EVIDENCE (all on v1.4.1 / main@6b2721a, same machine, minutes apart)

• Full suite (pytest tests): 32 failed — all under tests/analysis/java/ (test_java_sitter.py x15, test_inheritance_call_graph.py x8, test_java_analysis.py x5, test_jcodeanalyzer.py x3).
• Directory run (pytest tests/analysis/java): 11 failed.
• pytest tests/analysis/java/test_java_sitter.py in isolation: 21/21 PASS — and these are tree-sitter tests with no JDK dependency at all, yet 15 of them fail inside the full suite.

Something earlier in the suite corrupts shared state.

PRIME SUSPECTS

1. JCodeanalyzer._get_codeanalyzer_exec mutates GLOBAL process state: os.environ["JAVA_HOME"] = str(java_home) (cldk/analysis/java/codeanalyzer/codeanalyzer.py ~line 116) — leaks across tests.
2. eager=True tests wiping/rebuilding the shared .codeanalyzer cache dir mid-suite while other fixtures hold paths into it.
3. cwd or tree-sitter language-registry pollution from the python/typescript test families that run first in pytest tests ordering.

FIX DIRECTION

• Reproduce minimally with pytest -p no:randomly --lf-style bisection or pytest tests/analysis/python tests/analysis/java/test_java_sitter.py pair-wise to find the polluting family.
• Remove the global env mutation (pass env to subprocess.run instead of os.environ); isolate cache dirs per test via tmp_path fixtures.
• Gate: full-suite java results == directory-run results == per-file results.

BLOCKS a meaningful CI-green (#255) — the release-gate flip cannot land while the suite's failure count depends on invocation shape. Sibling issues from the same investigation: #256 (deterministic source-code crash), #257 (TS fixture gating).