docs: the slot-scopes number bounds half the question, not all of it

[fadion]

#74 said the deep-versus-shallow name lookup difference was "the ceiling on what slots could win". It is not, and that is the kind of overclaim that gets read later as a settled answer.

Both benchmarks pay exactly one successful map lookup. What differs is the failed lookups at the scopes in between, so the fifth-more-time figure bounds chain-walking and says nothing about the other half of the change — replacing that final map lookup with an array index. No benchmark in the file can see that half, and none was built to.

What changed

• Rewrote the Known Gaps entry to say which half is measured and which is not.
• Recorded that the cheaper half stands alone: Ref.Hops already says how far up a binding lives, so the evaluator could walk exactly that many parents and do one map lookup instead of up to Hops + 1. That collects the entire measured difference while vars stays a map, with no scope sizes to plumb and no change to define or assign.
• Recorded why either version is bigger than swapping a data structure: the evaluator has no runtime dependency on resolution at all — i.info is assigned in two places and read in none. The paths that would still need names are real: the REPL's :vars walks globals.vars, an aliased import builds its module from scope.vars[name], and four New(file, nil) call sites have no Info to consult.
• Tightened the same loose claim in the benchmark's own comment, which is where somebody actually reads it.

No behaviour changed.