Test Quality

Use when adding or changing automated tests.

Gates

Test one behavior contract, not incidental implementation details.
Use deterministic fixtures from data/mods/TEST_DATA for data with random, balance, or evolving behavior.
Avoid assertions that rely on probability, wall-clock time, locale, file order, or current global options unless the test pins them.
Reproduce reported flakes with the failing --rng-seed before accepting the fix.
When CI test jobs fail, inspect the job log for the exact test command, filters, shard split, user-dir, order-sensitive included tests, --rng-seed, and backend flags. Run that same shard command on the final rebased HEAD before declaring the fix validated; do not substitute a narrower tag filter or a different full-test command.
Before pushing test fixes that touch global state, NPCs, maps, overmaps, player position, RNG, options, or fixtures, run both the relevant focused test and the CI-equivalent shard/full command that can expose state bleed from earlier tests.
Keep Lua/C++ integration tests on real bound objects when the binding behavior is under test.
Expected DebugLog(DL::Error, ...) still makes the test process fail even when assertions pass. Use capture_debug_errors_during for the injected failure only and assert its exact messages; never clear the global observed-error flag or disable the test runner's error check.

Restore global state with restore_on_out_of_scope for calendar, options, globals, and singleton state.
Use clear_map() and explicit player placement when map contents matter.
Do not leave creatures, items, hooks, worlds, temp files, or option changes that can affect later tests.
Prefer new TEST_DATA fixtures over mutating base-game data in the test body.