perf: benchmarks, and a number for the slot-scopes question

[fadion]

The characterization suite is the arbiter of meaning and nothing played the same role for cost. Aria was never trying to be fast, and docs/compatibility.md puts performance outside what a version promises, so these are not targets to hit — they are for noticing when a change costs something nobody meant to spend.

What changed

• internal/interp/bench_test.go: Go's testing.B rather than a harness of our own, so benchstat, -count and -benchmem come free.
• Three phases separately — parse, resolve, eval — because a regression should point at a stage rather than at "the number moved".
• Macro workloads are examples, not fixtures written for the benchmark, so check-examples.sh already stops them rotting into something that measures nothing.
• Micro workloads for what the examples mix together: name lookup at depth, call overhead, and building an immutable collection.
• CI runs them at -benchtime=1x, which measures nothing and proves they still run.

The standard library had to be handled first

Eval loads it on every call: 1.3ms and 11,005 allocations, against 308 for an empty program. A workload measured through Eval is 97% stdlib loading. So it gets its own benchmark, being a real cost every aria run pays, and everything else is built with the timer stopped.

I checked that StopTimer excludes allocations and not merely time — a trivial program through the same path reports 0 allocs/op, so the numbers are the workload rather than the setup.

Read allocs/op, not ns/op

Every collection operation returns a new value, so allocation is the dominant cost, and a deterministic program allocates identically on a laptop and a shared runner. Wall-clock does not, which is why nothing gates on it — that would need a stored baseline that goes stale and cries wolf.

The last Known Gap now has evidence

name-lookup-deep and name-lookup-shallow run the same loop and the same arithmetic, differing only in whether the names are three scopes up or beside the loop, so the gap is chain-walking alone: about a fifth more time and ten more allocations out of eleven thousand. Pointer chasing, not garbage. That is the ceiling on what slot indices could win, in a case built to make the chain as expensive as possible.