Fix and simplify gentest randomisation toggle

The comment correctly described the behaviour we wanted, but the code
logic was backwards. This is easy enough to do when you've got an env
var called "RANDOMIZE" toggling a config value called "derandomize" so
we switch the name of the env var to match.

This makes it clear that there's exactly one circumstance where we don't
want randomisation and that's in the CI tests.

Related Slack thread:
https://bennettoxford.slack.com/archives/C069YDR4NCA/p1739440049238879

.github/workflows/generative-tests.yml

@@ -28,7 +28,6 @@ jobs:
 
           export \
             GENTEST_EXAMPLES=150 \
-            GENTEST_RANDOMIZE=t \
             GENTEST_MAX_DEPTH=25
 
           # We get a maximum of 6 hours runtime with Github Actions so breaking

justfile

@@ -219,7 +219,6 @@ test-unit *ARGS: devenv
 # Run generative tests using more than the small deterministic set of examples used in CI
 test-generative *ARGS: devenv
     GENTEST_EXAMPLES=${GENTEST_EXAMPLES:-200} \
-    GENTEST_RANDOMIZE=${GENTEST_RANDOMIZE:-t} \
       $BIN/python -m pytest tests/generative "$@"
 
 
@@ -228,6 +227,7 @@ test-generative *ARGS: devenv
     #!/usr/bin/env bash
     set -euo pipefail
 
+    GENTEST_DERANDOMIZE=t \
     GENTEST_EXAMPLES=${GENTEST_EXAMPLES:-100} \
     GENTEST_CHECK_IGNORED_ERRORS=t \
       $BIN/python -m pytest \

tests/generative/test_query_model.py

@@ -84,14 +84,7 @@
 settings = dict(
     max_examples=(int(os.environ.get("GENTEST_EXAMPLES", 10))),
     deadline=None,
-    # On CI we want to run derandomized unless we're explicitly in the
-    # long-running bug finding tests. This prevents flaky CI.
-    # In development we never want to run with derandomize because it's
-    # less effective at finding bugs and, more importantly, has very slow
-    # replay due to turning off the test database
-    derandomize=bool(os.environ.get("GENTEST_RANDOMIZE"))
-    if os.environ.get("CI")
-    else False,
+    derandomize=bool(os.environ.get("GENTEST_DERANDOMIZE")),
     # The explain phase is comparatively expensive here given how
     # costly data generation is for our tests here, so we turn it
     # off by default.