## TL;DR
The fix passed in isolation but the full suite failed because a module-level cache leaked state between tests. Find the cache (module globals, singletons, lru_cache), clear it in setup or teardown, and never accept a healing patch without running the affected suite subset, not just the single test.

## The symptom
```text
agent's fix passed in isolation but failed in the full suite  -  a module-level cache leaked between tests
```

## Steps
### 1. Reproduce the suite-order failure
Run the healed test together with the tests that run before it in the suite (or the full file). Confirm it fails in company but passes alone.
Expected: Isolated pass, suite failure: order-dependent state confirmed.

### 2. Find the leaked cache
Look for module-level dicts, singletons, functools.lru_cache, and class-level registries in the code under test. Add a debug print of cache size in setup to catch the leak red-handed.
Expected: The specific cache object identified, holding entries from earlier tests.

### 3. Clear it in setup/teardown
Add cache clearing to the test setup (or an autouse fixture): call cache_clear(), reset the singleton, or reimport the module. The fix should make the cache lifecycle explicit.
Expected: The healed test passes both alone and in the full suite.

### 4. Require subset runs for every heal
Change the healer's acceptance gate: a fix must pass the target test plus its file (or the tests sharing the module) before it is accepted. Single-test green is not sufficient.
Expected: No heal is accepted on isolated evidence alone.

## When to use
- Healed test passes alone but fails in the suite
- Failures depend on which tests ran earlier
- Module-level caches, singletons, or registries exist in the code under test

## When not to use
- The test fails in isolation too (different problem)
- The ordering issue is in the test code itself (fix the test's assumptions)
- The suite is intentionally order-dependent (rare; document it)

## Why it happens
Module-level state persists for the whole test process. In isolation the cache starts empty and the test passes; in the suite, earlier tests populate it and the healed test sees stale entries. The agent validated against the easy case and missed the real one.

## Edge cases
- lru_cache on methods: clear via the underlying function attribute.
- Caches populated at import time: reimporting the module in setup may be needed.
- Thread-local caches: clear per test, not just per process.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_o-PEYHWg8T8r9KJKg7Ze5g