Fixture consultation — dl, round 1 (synthetic)

A synthetic consultation answer for this red-team corpus. The speaker is the
expert scope `dl`. Where it quotes or names a third party, that party is the source.

On the ceiling: the admissible band is between 15M and 20M edges, scaling linearly at
about 0.51 GiB per million edges at 128 dimensions. Fixture paper B (Okafor and
Lindqvist) shows that the expander bound tightens under mean aggregation; that is their
theorem, not mine. I label the saturation figure inferred from the two tables, not
measured. Poursafaei et al. did not land in anything I retrieved. A non-inferiority
claim needs a margin stated before the data is seen, otherwise it can be met by any
result. A metric on which no-refresh wins measures stability, not faithfulness; it
must be reported as such, never as faithfulness. My own expectation is that under mean
aggregation the stale fraction, not the count, governs the error. Their bound is tight
only under mean aggregation.
