Benchmarks: Time::Keeper::Natural against the marks of
plan_port_natural_xs/13-benchmarks-and-the-gate.md, set before the port was
measured (the pure-Perl 0.01's were in plan_time_keeper_natural/11). There is no rival to compare with: no module on CPAN reads these
phrases in these locales, so each mark is a rate or a size of its own.

From the dist's top, after `perl Makefile.PL && make`, with Time::Keeper 0.07
or later installed or on PERL5LIB:

    perl -Mblib bench/run.pl            # every mark
    perl -Mblib bench/03-tomorrow.pl    # one

A rate is calls a second by Benchmark's countit over half a second of CPU, the
best of five; Benchmark subtracts its own empty-loop overhead, which flatters a
fast call, so each line also gives the plain wall-clock rate. Every rate is
the same text parsed again and again, and nothing is cached between calls, so
repetition does not flatter it.

    TKN_BENCH_SECONDS   seconds a run (0.5)
    TKN_BENCH_RUNS      runs, the best kept (5)
    TKN_BENCH_LOADS     fresh processes for the load time (41)

Benchmarks 10 to 16 hold the marks of plan_time_keeper_natural_2/13 (0.02 and
later), also set before any measurement: intervals, day ordinals, a zone in
the text, fractions, a duration after a point, find, and the first use of a
locale with its zones.

Measured 19 October 2026 (Time::Keeper 0.07) and 20 October 2026 (0.04 on
Time::Keeper 0.09) on an Apple M5, perl 5.42: see the numbers in
plan_port_natural_xs/13-benchmarks-and-the-gate.md and
plan_time_keeper_natural_2/13-benchmarks-and-the-gate.md.
