docs(readme): say where the map pays and where it does not, before someone measures it for us

The page claimed a token saving and never said when it fails to hold. A map is a fixed cost paid
once per context, so it pays in proportion to what that context goes on to do -- strongest when a
lane starts cold on a tree it does not know (an orchestrator matching tasks to models spawns many
such lanes), strongest again on the checking pass, and NEGATIVE on a lookup one grep already answers.

That last case is not hypothetical. A third-party A/B on six read-only localization tasks
(nikzlabs/shipit#2684) measured ripwire at 83.5% of baseline overall but MORE expensive on three of
the six, the losses concentrated where a single grep sufficed. They used the CLI, the documented
verb, and our own skills' instruction to run it first. The number is theirs and it is fair; the
missing sentence was ours.

The multi-agent claim is labelled as reasoning, not result: the published figures are single-agent,
the scaling hypothesis is pre-registered and unrun, and the page now says so rather than implying a
measurement that does not exist.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>