[tannevaled]

Closes the first two parts of #747 and gives the third one a guard.

The clock

bench/modules/run.sh timed with /usr/bin/time -p, which prints two decimals
of a second:

$ for i in $(seq 10); do { /usr/bin/time -p /usr/bin/true >/dev/null; } 2>&1 | awk '/real/{printf "%s ",$2}'; done
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

10 ms resolution, and anything under one tick reads zero. Of the 171 timing
cells in the three per-module sections of BENCHMARKS.md, all 171 are
multiples of 10 ms; 36 are ten ticks or fewer.

That is where the bolded rows lived. matrix | 20ms | 50ms | 0.40× was a
two-tick measurement. tsort | 80ms | 90ms | 0.89× claimed an 11% win over
MRI from a difference of one tick — re-taken, it is 1.42×, a loss.

Three changes to the harness

• Time::HiRes instead of /usr/bin/time -p, so the figures carry the
  precision the ratios were already quoted to.
• A run that exits non-zero is an error, not a fast run. best() kept the
  smallest time of N without looking at the exit status, so a command that died
  immediately won its own series and got published. It now marks that arm and
  carries on with the others.
• The runtimes are interleaved, one run each in rotation with the order
  rotating every round, instead of each taking its own block of wall-clock time.
  A machine whose load moves during a block charges the movement to whichever
  runtime was running, and nothing in the output says so. See Compare implementations, not moments: the Fugue row was 14% too high go-crdt/crdt#131:
  the same defect put a published ratio 14% out, and contention does not fall on
  every arm alike — it falls hardest on the shortest-running one, which is
  the fastest one.

The table, re-taken

A new dated section; the June ones are kept as they were taken. All 35 modules
passed the byte-identical-to-MRI gate under every installed runtime, no row was
dropped and no run exited non-zero.

MRI is the control. Across the 24 rows the two runs share it reproduces
within a few percent — 220→215 ms on set, 1450→1546 on did-you-mean,
570→567 on scanf, 330→329 on securerandom — so the machine and the workloads
are the same, and what moved in the rbgo column is rbgo's.

Five of those 24 rows change which side of parity they sit on:

	June	now
cmath	0.67×	1.20×	a win that is a loss
prettyprint	0.89×	1.42×	a win that is a loss
tsort	0.89×	1.42×	a win that is a loss
prime	1.67×	0.94×	a loss that is a win
ipaddr	1.91×	0.54×	a loss that is a win

Three of the eight modules the June text lists as "rbgo beats MRI outright" are
not beating MRI, and all three sat on one or two ticks.

Other rows moved because the code did, which is the larger part: strscan
10720 → 262 ms and json 1520 → 701 ms are #89 and #88 landing the day after
those tables were taken. regexp 1530 → 4968 ms moved that far the other way.
Three months separate the runs; the new section does not try to attribute each
row, only to say which comparisons the old clock could not support.

set, and the guard

set reads 38.2× where June published 10.09×, and that one is not the clock
either. #56 bound Set to go-ruby-set/set on 2026-06-29; #330 reimplemented it
here on 2026-08-08 for MRI 4.0 conformance, with the bulk of the Ruby-observable
API in the prelude; #333's go mod tidy dropped the now-unused require the next
day. The benchmark exercises exactly that algebra, so it now measures prelude Ruby
where it used to measure Go.

Nothing about that trade was wrong — conformance was the point. What was missing
was anything that would say it had happened, and a page in another org went on
attributing this number to a library we had stopped using, for seven weeks, with
a figure that had moved 3.7× underneath it.

bench/modules/BACKING.md now records which modules run on an external
go-ruby-<mod> library and which run on code in this repository. It is generated,
and TestBenchmarkedModulesSayWhereTheirCodeLives fails when it and go.mod
disagree — I checked that it does, both by flipping one row back to
github.com/go-ruby-set/set (it names the line) and by deleting the file. Today
complex, rational and set are in-tree.

The four documentation pages in the other orgs are untouched; #747 records which
they are.

🤖 Generated with Claude Code