AP-6 Canonical tasks benchmarked against the generic path

A product whose surface agents consume SHOULD benchmark named canonical tasks through that surface and through the generic interface the same tasks would otherwise use, with a gate that fails when a task's calls or tokens pass their ceiling or its ratio to the generic path falls below its floor.

Field	Value
Keyword	SHOULD
Level	profile
Checked	by judgement
Findings	F409
Harness facts	none
Threat	construct: each baseline was written by the author of the surface it is compared with
Status	active

Rationale: AP-1 caps each surface, but a task that got dearer by taking more calls, each under its ceiling, passes every one of them. Shio and freewilly each ran named tasks both ways and held the result by a test (F409, R3/S4), and freewilly's assertion "caught it on the commit" [freewilly@c1c2eaf:agent-budget.json#L113-L114] that widened one response by 43 tokens. The generic path is the point of the comparison: Shio found that a single number per task "could no longer distinguish" [shio@821f18d74:docs/CHANGELOG.md#L478] a cheaper protocol from a wasteful console, and that its token advantage fell by more than half when its own REST reads were fixed. So a ratio is reported with its baseline, and its floor drops with a better baseline rather than being defended. What the benchmark measures is the shape of the path, not its transport, since Shio's shaped surface is HTTP as well. Both projects learned to isolate the inputs, since shared fixtures and a date in a fixture turned the gate red with no regression to blame, and neither asserts wall clock. The rule is a SHOULD although F409 admits MUST: roadkeep and winwright ship surfaces to agents with no such benchmark and no failure recorded against its absence, and building one costs a fixture for the generic path as well as the shaped one.