Lab analysis: median_steps_per_operation mixes assignment grain into an operation-grain metric (AM-U-007)

[roy-tong]

Found by @jaden3824 (Urusilla) in the second external conformance vector (urusilla-002, pinned urusilla @ cfb4ef3): with two operations in one assignment whose per-operation step totals are [3, 2], median_steps_per_operation reported 5.0 with denominator 1.

Root cause (lab/agentmeasure_lab/analysis.py): attempt steps were accumulated into steps_by_assignment and the median taken across assignments, while the metric name and label promise operation grain. A single assignment holding N operations therefore reports the assignment total as if it were one operation.

Fix (75027ff): steps are summed per operation key; the median runs over resolved (declared and reconciliation-passed) operations only; the denominator is the number of contributing operations. Fail-closed: unresolved operations and operations that failed #9 reconciliation contribute nothing. In the vector this yields 2.5 over 2; single-operation behavior (urusilla-001) is unchanged.

Regression guard: conformance/runners/run_external_fixture_002.py (CI) pins the per-operation totals [3, 2], the 2.5/2 median, full metrics reproduction against the externally pinned expected.json, and the fail-closed path (a tampered declaration removes that operation's steps from the median).

Same handling as #8/#9: externally found, fixed with the reporter's fixture as the regression vector. The vector is ingested at conformance/vectors/external/urusilla-002/ with the source claim boundary stated verbatim.

[roy-tong]

Ingested: conformance/vectors/external/urusilla-002/ (events, expected, and the vector report verbatim) with a new CI guard run_external_fixture_002.py that reproduces your pinned values on every run.

The grain fix is in 75027ff (reference corrected in 9997f2f): per-operation step totals [3, 2], median 2.5 over 2, and operations that fail reconciliation are excluded from the median (fail-closed). Vector 002 is now the regression vector for this defect, the same role 001 played for #8/#9.

Thank you again; three externally found defects in two vectors is exactly the conformance loop working.