The production-cluster Skiff run the format comparison says is owed

[sshaplygin]

Part of #63.

docs/format-comparison.md ran, reported, and was explicit about what it could not settle. The whole-map ratio between the Skiff leg and the typed YSON leg came out around 1.10–1.20×, consistent with the size of what goes down the output pipe rather than with decoding — and the local cluster cannot adjudicate it: rounds scatter about 2×, roughly half the measured wall time is process start and waiting for the first batch, and a local cluster reports no job CPU at all, which is the denominator the Skiff decision criterion is actually stated over.

So the run is owed on a cluster where the noise is smaller and job CPU exists. The harness is in-tree and the access exists; this is a matter of running it.

• run the existing harness on a production cluster, rounds interleaved and paired as before
• pre-register the prediction first, as the local run did
• report a spread from repeated runs, not a third single number — AGENTS.md says exactly this and it is the easy mistake here
• write the result into docs/format-comparison.md and docs/benchmarking.md whichever way it comes out, including if it takes the Skiff advantage away entirely
• keep it clear that this measures one task's map, and does not close docs/skiff-compatibility.md required test 4