Forcing scenario: ND ANN with budget exhaustion and replay divergence¶
Steps (MUST):
- Execute a NON_DETERMINISTIC ANN request with a strict budget (
max_vectors=0, low latency). - Allow partial execution; status MUST be
PARTIAL. - Fallback deterministic path MUST be used when ANN cannot proceed under budget.
- Replay MUST re-run the plan and detect divergence against the baseline fingerprint.
Invariants (IDs):
- INV-020: Randomness required for ND execution.
- INV-021: Budget exhaustion recorded, status PARTIAL.
- INV-010: Contract alignment enforced.
- INV-030: Plan fingerprint must not mutate between execution and replay.
Expected behavior:
- First run yields partial results and signature.
- Replay with the baseline fingerprint produces a different results fingerprint and details indicating nondeterministic divergence.
- Fallback path is exercised during replay to make divergence observable.