[kacy]

Closure calls and constant division carried checks that could only pass:

• A closure body reading its own capture with a constant slot now does one load. The closure is the body's first parameter, and every call through a closure value tests the handle before it jumps in, so the second handle test and the spill check could not fail. In make_adder's lambda the body went from 26 instructions to 10.
• A closure's handle test is reused for the rest of a straight run, so f(f(x)) tests f once. The cache is keyed by the local that holds the closure, dropped when that local is stored to, and cleared at every label.
• Storing a capture into a closure built on this function's stack, at a constant slot, is one store.
• A constant divisor other than 0 and -1 no longer emits divide-by-zero and overflow checks. Their conditions folded to constants, but the branches stayed: two dead tests on every i % 7.

Measured against main (t2.sh, 5 interleaved rounds, median ms):

bench	main	this	change
closure_calls	114	82	-28%
closure_error	288	273	-5%
http_head_read	646	614	-5%
the rest of the set			within noise

Under callgrind, closure_calls goes from 218M to 140M instructions and closure_error from 892M to 773M. A call through a closure now costs the same as a direct call in closure_calls.

Building maps and sets in place, which was planned for this change, is left out: it cost event_ledger 8% in timing with no gain anywhere, even though it ran fewer instructions.

The local suite passes, including both leak gates and memcheck.

Refs #986