capabilities: the broker answers a harness's initialize asking another MCP version with its own 2025-06-18, as MCP's version negotiation says (decision 0065 slice two, U9b1c) - #940

Hosted merge-queue runners vary in speed. The engine job's Linux leg ran
57 minutes on one pass of #940 and was cancelled at its 75-minute limit
on the next, mid capability_broker, with every shard 15-30% slower. The
coverage job runs 46-66 minutes and was cancelled at 75 on #933 and
#935. A cancelled job removes the group from the queue. Both jobs now
hold timeout-minutes: 90, under the queue's own check timeout, raised to
100 the same day. contributing.rs pins the two lines.

Ruling: operator, 2026-11-01, "A": job limits 90, queue 100.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Signed-off-by: Valentin Yanakiev <valentin.yanakiev@gmail.com>

One runner ran every shard one after another: a hosted merge-queue run
took 57 minutes on Linux and 42 on macOS, and past 75 on a slow runner,
which dropped #940 from the queue. scripts/test-shards.sh now names each
shard's machine in each system's layout, and a run whose
BROKKR_TEST_MACHINE names one (linux/<n>, macos/<n>) runs that machine's
shards alone, still one after another, so a shard never shares its
runner. Unset, every shard runs, as a local run does. The variable is
read once and unset, so no test a shard runs inherits it. Each system's
machines are numbered from 1 with none empty, and a machine named any
other way is refused before anything is built. `machines test|coverage`
prints the layouts.

Linux: libraries | the it binaries, heaps, other own-process binaries
and doctests | capability_broker alone; each about 11, 22 and 20 minutes
on a hosted runner. macOS: everything but | the other own-process
binaries; about 21 and 19.

ci.yml's engine job becomes engine-machines, one leg per machine (45
minutes each), and `engine` is the verdict over them under the names main
requires, `test (ubuntu-latest)` and `test (macos-latest)`. It runs
`always()`, since a skipped job reports success to a required check: a
machine that failed, was cancelled or never ran fails both verdicts. The
macOS startup gate runs once, on macos/1.

contributing.rs holds a required job's needed job too (JOB_LINES), reads
a matrix's included systems, and gives a verdict's leg its machines'
steps, so the guide's rows still write a local run of the whole suite.
test_shards.rs holds the selection, the refusals and that ci.yml's
matrix is the layouts. Removal controls: dropping the unset fails the
machine test (a shard inherited it); dropping the empty-machine check
fails the layout test.

Ruling: operator, 2026-11-01, "yes, do both": the slow checks split across parallel hosted machines.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Signed-off-by: Valentin Yanakiev <valentin.yanakiev@gmail.com>