[johng]

Started from "why are tests skipped on CI -- flaky perf work on shared boxes,
or real issues?" It is neither: the skips are mostly legitimate platform
gates, but CI had no way to tell those apart from tests that were not running
for fixable reasons, and several were not running at all.

What was wrong

Files that execute zero tests reported PASS. A file where every
collected test skips exits 0, so classify() called it a pass and the tally
counted it as one. _summary_line() did not match "skipped", so
"10 skipped in 0.70s" matched nothing and the file printed with a blank
summary -- indistinguishable from one that ran. On run 35022043268, green
throughout:

leg	files that executed no tests
ubuntu / 3.14	14 (reported as "236 passed, 0 skipped")
ubuntu / 3.13	16
macOS / 3.14	26

tests/test_greenlet_interop.py has never run on CI. It is in that set on
every leg. CLAUDE.md names it the guard for the greenlet/TLBC coexistence
invariant. Nothing installs greenlet, so all 10 of its tests skip -- and it
looks green. Verified: install greenlet (3.5.6, cp314t wheel) and it is
10 passed on 3.14t.

tests/test_chan_properties.py is skipped on both 3.13 legs for a fixable
reason. The comment blamed PyO3's lack of free-threaded support below 3.14 --
true, but only of the newest hypothesis. 6.156.1 is the first release with a
Rust core (the first with platform wheels instead of py3-none-any); 6.155.7
and earlier are pure Python. Verified: with hypothesis<6.156 the file is
4 passed on 3.13t (4 property tests x 200 examples over channel FIFO /
conservation / close / select). Same shape for cryptography + trustme, which
gate 5 TLS tests and do publish cp314t wheels.

The contention valve was inverted. 38 pytest.skip("... timed out (shared-box contention)") sites across 22 files turn a hung child into a
silent skip. Their deadlines were literals -- the only deadlines in the suite
ignoring RUNLOOM_TIMEOUT_MULT, which CI sets to 2 precisely because its
runners are slow. So the scaler doubled the outer per-file ceiling and left
the inner ones fixed: on the box the scaler exists for, the inner one trips
first and the test skips instead of running slower. I found no evidence of it
firing in the runs I sampled -- this closes the mechanism, not an observed
failure.

What this changes

• tests/run_isolated.py -- ALLSKIP, the missing third verdict; -rs so
  skip reasons reach the output; two unconditional end-of-run blocks (the
  files that executed nothing, and a census of skipped tests by reason); the
  tally now reads N passed, M ran nothing, K not-passed.
• tools/ci/test_patched_cpython.sh -- best-effort install of
  greenlet / hypothesis / cryptography / trustme, one pip call per package
  (a combined install is one transaction, so one unbuildable package takes the
  whole set down -- exactly what happens on 3.13t); hypothesis fallback pin;
  and the step summary now reports coverage instead of just "passed".
• tests/adv_util.py + 22 files -- child_timeout(), applied at exactly one
  point per call chain so the scaling cannot compound. No behaviour change at
  the default mult of 1.

Verification

Full suite green on 3.14t: 204 passed, 32 ran nothing, 0 not-passed.
Exit-code propagation through the new tee pipeline checked for
child-failure, child-success and cd-failure. Every claim above about a
package installing or a test passing was measured, not inferred.

Two findings this PR deliberately does NOT act on

1. The REALTHREAD gate is over-broad, and hides a real bug. The largest
single gap in the suite is 245 skips from one gate -- the stdlib conformance
tests that drive cooperative primitives from real OS threads, gated off as
deadlocking. Probed on 3.14t:

file	with RUNLOOM_STDLIB_REALTHREAD=1
test_stdlib_queue_monkey.py	76 passed, 3 skipped -- no deadlock
test_stdlib_socket_monkey.py	79 passed, 8 skipped, 3 failed
test_stdlib_threading_monkey.py	TIMEOUT at 240s in TestRunloomCondition::test_acquire

The gate is correct for threading and only for threading; it is costing
~155 conformance tests to protect against a deadlock in one file. Narrowing it
wants a Linux run and repeats first -- a deadlock in CI costs a 35-minute job
timeout.

2. Scoped IPv6 literals lose their scope id under monkey.patch(). Two of
those three socket failures are one bug. They pass on stock CPython 3.14t
and fail only under monkey:

stock  : ('fe80::1', 1234, 0, 1)
monkey : ('fe80::1%lo0', 1234, 0, 0)

src/runloom/monkey/dns.py:435, in _patched_getaddrinfo:

if aaf == socket.AF_INET6:
    sa = (a, port, 0, 0)

The address goes through verbatim with scope_id hardcoded to 0, instead of
splitting addr%iface and resolving it via if_nametoindex (it also skips
stock's inet_ntop canonicalisation -- hence 8D vs 8d in the assertion).
Not cosmetic: a link-local IPv6 address cannot be connected to without a scope
id, so connect() / create_connection() to fe80::...%eth0 works on stock
and fails under monkey. The third failure is a smaller divergence in the same
layer -- _patched_getnameinfo routes bad input through _blocking_call's
thread pool and CPython's testRefCountGetNameInfo, which catches
TypeError, sees a different type come back.

Both belong in their own change with their own tests.