improve TESTS_ORACLE_PROMPT.md with semantic impact analysis

Author: rnetser

TESTS_ORACLE_PROMPT.md

@@ -78,16 +78,51 @@ RIGHT: "Smoke test X uses fixture Y, which depends on the changed fixture Z"
 
 ### For This Repository
 
-This is a **test suite repository** — the changed files ARE often the tests themselves, or test infrastructure (fixtures, utilities, conftest files). Adapt analysis accordingly:
-
-- If changed files are test files: recommend running those changed tests, plus any other tests that share fixtures, utilities, or base classes with them.
-- If changed files are test utilities/fixtures/conftest: recommend running all tests that depend on or import from the changed utilities.
-- If changed files are under `utilities/` or `libs/`: trace which tests import from these modules and recommend those tests.
+This is a **test suite repository** — the changed files ARE often the tests themselves, or test infrastructure (fixtures, utilities, conftest files). Adapt analysis accordingly.
+
+### Semantic Impact Analysis (CRITICAL)
+
+Do NOT recommend tests based solely on import relationships or file proximity. Analyze the **semantic impact** of each change:
+
+**Changes that DO require test execution:**
+- Modified test logic, assertions, or parametrization
+- Modified fixture behavior (return values, setup/teardown, scope changes)
+- Modified utility functions that are called during test execution
+- Changed function signatures that tests depend on
+- Bug fixes in shared code paths
+
+**Changes that do NOT require test execution:**
+- New opt-in features behind CLI flags (e.g., `--analyze-with-ai`) that are disabled by default
+- New imports in conftest.py that only activate under explicit opt-in conditions
+- Documentation files (`.md`, `README`)
+- Linter configuration (`.flake8`, `.ruff.toml`)
+- Dependency additions for opt-in features
+- Session lifecycle hooks (`pytest_sessionstart`, `pytest_sessionfinish`) that do not affect test collection, execution, or results under default settings
+- Coverage configuration changes
+
+**For conftest.py changes specifically:**
+1. Check if the change affects test COLLECTION (markers, parametrization, item modification)
+2. Check if the change affects test EXECUTION (fixtures used by tests, hooks that modify test behavior)
+3. Check if the change is INERT under default settings (opt-in flags, conditional imports)
+4. Only recommend tests if the change affects collection or execution under default settings
 
 ### Dependency Tracing
 
-1. Examine code changes in each modified file
-2. Identify affected code paths, functions, and classes
-3. Analyze pytest-specific elements: fixtures (scope, dependencies), parametrization, markers, conftest changes
-4. Trace test dependencies through imports, shared utilities, and fixture inheritance
-5. Detect new tests introduced in the PR
+When changes DO require test execution, trace dependencies precisely:
+
+1. Examine code changes semantically — understand WHAT changed, not just WHERE
+2. Identify affected code paths and whether they execute during normal test runs
+3. Analyze pytest-specific elements: fixtures (scope, dependencies), parametrization, markers
+4. Trace ONLY active dependency paths — imports that are actually called during test execution
+5. Distinguish between direct dependencies (test calls changed function) and inert dependencies (import exists but code path is gated behind a flag)
+6. Detect new tests introduced in the PR
+
+### Verification Before Recommending
+
+Before recommending any test, answer these questions:
+1. Does this test actually CALL or DEPEND ON the changed code during a normal test run?
+2. Is the changed code path active by default, or gated behind an opt-in flag?
+3. Would the test behave DIFFERENTLY after this change under default settings?
+4. Can I trace a concrete execution path from the test to the changed code?
+
+If the answer to questions 1, 3, or 4 is "no", do NOT recommend the test.