Repository navigation
TEST 120
Inject synthetic refusal/deadline/cancellation at input and model/checkpoint preflight, identity preflight, metadata, inference, identity postflight and finalization boundaries. Advance fake monotonic time and count requests/retries, including failure before I/O, during an in-flight call, after a completed response, and during retry backoff. Place fake secret/path/prompt strings in model responses and exception messages. Request a private observation, then try unsafe permissions, symlink/occupied destinations, partial JSON, substituted successful schemas, and writer/finalization failure. Run both review adapters and their CLIs without real inference. A separate live approval may authorize one diagnostic invocation with its own immutable bounds; tests never grant that authority.
Only closed host-authored enums, approved limits, counters, elapsed time, in-flight uncertainty and hashes appear in the distinct non-success observation. Canary/path/raw model or exception text is absent from observations, logs and finalization diagnostics. Private custody and atomic non-overwriting output hold; unsafe or unsupported targets create no success-shaped fallback. The original error/cancellation category and exit code remain unchanged even when observation generation or persistence fails. A successful run does not write a refusal record, and partial/refusal data cannot become a completed receipt. No new request, retry, deadline expansion, model/service mutation, or server-completion assertion is performed merely to populate diagnostics. Executed targeted regressions named in REQ-119 cover all four model request phases, before-I/O invalid input, after-response fake deadlines, transport canaries, reused-observer reset, CLI cancellation/finalization, unsafe/occupied/private atomic custody, writer failure preserving the primary exit, and success separation. Focused Python review/run/contract tests passed 204 cases and .NET GenerationTests passed 97; full suites passed 2,166 Python and 108 .NET tests. PR #247 passed all seven applicable gates, including Windows custody and Ubuntu execution, before integration as fbd69b0ae6480ae97b725c325b4ad05fe5189ba1. These executed cases do not assert exhaustive coverage of every combination in the stimulus design. The single separately authorized metadata-only L2 produced a private valid inference-deadline observation with exit 2, three requests, zero retries, possible in-flight state and unknown server completion, with no final receipt. That is diagnostic evidence, not a passing assessment, server-termination proof or native-build qualification; no retry followed.