Cross-language refusal tests read two different env var spellings: CLDK_TEST_PYTHON_NEO4J_URI vs CLDK_TEST_NEO4J_PYTHON_URI

[rahlk]

Problem

The same concept — "a live graph of another language, to prove this backend refuses it" — is spelled two ways:

suite	variable
tests/analysis/java/test_java_schema_probe.py	CLDK_TEST_PYTHON_NEO4J_URI / _USER / _PASSWORD
tests/analysis/typescript/test_typescript_refuses_python_graph_live.py	CLDK_TEST_NEO4J_PYTHON_URI / _USER / _PASSWORD

PYTHON_NEO4J in one, NEO4J_PYTHON in the other. Setting the wrong one is silent: the module skips, and a skip reads as a pass.

Why it matters more than a papercut

Found by running the Java suite with what I believed was every gate satisfied, seeing two skips remain, and only then noticing the transposition. The whole point of #360 was that a silently skipped module hid five legs of rot; a variable name that is nearly right is the same failure mode in miniature.

Scope boundary

In scope: one spelling, both suites, and the skip reason quoting the variable it actually reads.

Out of scope: the CLDK_TEST_NEO4J_* (read-only) and CLDK_TEST_NEO4J_WRITE_* (disposable) families, which are consistent and whose separation is deliberate (#324).

Goals

• Both suites read the same variable name
• Each skip reason names the exact variable, so a transposition is visible in the output rather than requiring a reader to compare two modules
• The new pytest_terminal_summary hook from Write-gated Neo4j parity modules have been skipped through every leg; all ten tests now fail against current behaviour #360 reports this gate too, alongside the write-gated one

Definition of done

• Setting one documented set of variables satisfies the cross-language refusal gate in every suite that has one.

[rahlk]

A third spelling in the same family, found while running the Java suite: the application name is CLDK_TEST_NEO4J_JAVA_APP for Java and CLDK_TEST_NEO4J_APP for TypeScript and Python.

Setting the TypeScript spelling against the Java suite is silent and looks like it worked: 563 passed / 75 skipped instead of 636 passed / 2 skipped. Nothing fails; 73 tests just quietly do not run. That is worse than the transposition this issue was opened for, because the run still ends green.

So the family is:

concept	Java	TypeScript / Python
application under test	CLDK_TEST_NEO4J_JAVA_APP	CLDK_TEST_NEO4J_APP
a graph of another language, to prove refusal	CLDK_TEST_PYTHON_NEO4J_URI	CLDK_TEST_NEO4J_PYTHON_URI

Both belong in the same fix, and the terminal-summary hook from #360 should report a suite that skipped for want of an application name, not only one that skipped for want of a write URI.

[rahlk]

Surveyed every CLDK_TEST_* variable the suites actually read. There are five inconsistencies, not the two this issue was opened for.

concept	Java	TypeScript / Python
application name	CLDK_TEST_NEO4J_JAVA_APP	CLDK_TEST_NEO4J_APP
cross-language graph URI	CLDK_TEST_PYTHON_NEO4J_URI	CLDK_TEST_NEO4J_PYTHON_URI
its user / password / app	CLDK_TEST_PYTHON_NEO4J_{USER,PASSWORD,APP}	CLDK_TEST_NEO4J_PYTHON_{USER,PASSWORD,APP}
scale fixture app	CLDK_TEST_NEO4J_JAVA_SCALE_APP	—
default when unset	none, so the suite skips	odoo-slim-19 / superset-frontend

The last row is its own footgun and is arguably worse than the transposition. Python and TypeScript default their application name, so pointing CLDK_TEST_NEO4J_URI at a graph that does not hold that application skips silently — the run ends green having verified nothing. Java has no default and skips loudly by comparison.

Proposed resolution

• One spelling per concept: CLDK_TEST_NEO4J_APP, CLDK_TEST_NEO4J_SCALE_APP, and the CLDK_TEST_NEO4J_PYTHON_* family for the cross-language refusal graph — keeping the CLDK_TEST_NEO4J_ prefix throughout so the family sorts together and a transposition is visually obvious.
• A retired spelling that is still set becomes an error, not a shrug. Anyone with CLDK_TEST_NEO4J_JAVA_APP in their shell history would otherwise get exactly the silent skip this issue is about. A few lines in tests/conftest.py that fail the run naming the old variable and its replacement costs nothing and removes the failure mode permanently.
• Every skip reason names the exact variable it looked for, and the terminal-summary hook from Write-gated Neo4j parity modules have been skipped through every leg; all ten tests now fail against current behaviour #360 reports a suite that skipped for a missing application name, not only one that skipped for a missing write URI.

Open question worth a ruling: whether the defaults should go away entirely. A default application name is convenient locally and is the reason a mis-pointed URI reads as success. My inclination is to drop them and require the name, since these suites already require a URI with no default for the same reason (#324).

Sequencing: this lands after #354, which is in flight and editing the same Java live-test files.