Coding fixture contamination — invalid evidence retained. Pi received an absolute @prompt file and could edit the source fixture directory. Subsequent coding runs therefore did not start identically. All eight coding runs are invalid, including apparent passes. The private harness now supplies prompts on stdin and copies only declared, hash-checked files. Original fixtures were restored; their tests fail as intended; repaired runner syntax passes. No live retest: the 24-run ceiling was consumed. Operational invariant: fixture sources must be inaccessible to the worker; verify sandbox input hashes before every run. Full coding coverage remains missing.
Reasoning ceiling — retain baseline. A exhausted 4,096 tokens twice. B completed once, then exhausted 16,384 tokens. No repeated rescue; no default-budget change. This is a bounded runtime/client observation, not proof of quantization damage.
Isolated router startup — corrected before smokes. Copying unrelated media declarations required an unavailable media endpoint reference. Temporary test configurations were narrowed to the explicit primary route and its admission/auth policies. Both routes then passed authenticated discovery, unauthorized 401 and cap-clamp checks. No production file changed.
Publication worker unavailable — local completion. No independent reviewer output was produced. Deterministic grading and evidence consistency checks remain independent of the evaluated GLM; no promotion review is claimed.