Release plan: correct the intl/gmp rates for the ignored SKIPIF guards

[PurHur]

The Phase 1 numbers count cases the corpus itself declares inapplicable.

--SKIPIF-- was parsed into $sections['SKIPIF'] by test/BaseTest.php and then never read — and it was not in UNSUPPORTED_SECTIONS either, so it did not throw. Every skip guard in the corpus was dead (#24888, fixed by #24897).

Measured at cc50fbb8f

cases carrying --SKIPIF--	671
of those, executed	455
of those, failing in the committed baseline	81
of those 81, guard confirmed to evaluate to skip	≥ 19

no MYSQLI_TEST_DSN, no pg_connect (libpq FFI), requires PHP 8.4+, inotify unavailable — not compiler defects.

Effect on the #24697 scoping decision

Of 56 failing intl cases, 46 carry an ignored guard, but only 2 are confirmed skips. The other 44 were inconclusive under a standalone evaluator that runs the block outside the harness env — recorded as inconclusive, not as "would run".

So the honest intl number may fall by anywhere between 2 and 46, and the 88.9 % rate cannot be relied on until a collect is taken with SKIPIF honoured. Scoping intl out, or committing to fix 56 cases, would both be priced off a contaminated figure.

gmp is unaffected: 6 failing, none of them skip cases.

Also sharpens "the pattern worth carrying forward"

Three failures of the same shape landed today, two of them in this plan's own tooling:

• the dead --SKIPIF-- — 671 cases stated when they were inapplicable and the harness discarded every statement, while ~1,100 lines of hand-written str_contains($name, …) rules grew in VMTest.php to do the same job
• the quarantine that hid 6 real failures by comparing two collects 15 commits apart (Compliance baseline is environment- and commit-specific: CI diff against a dev-box baseline conflates regressions with host differences #24726); re-measured against a pinned pair, 8,112 executed cases produced exactly one genuine flip
• a baseline of failing names that could not tell "fixed" from "no longer run" — the nightly reported four non-executing cases as FIXED, and the prescribed response to FIXED is "regenerate the baseline" (CI: select the compliance suite; add a spine-chunk split-TU probe #24835)

The through-line is sharper than the original wording: it is not enough for a check to run. It must be impossible to satisfy by absence. A case vanishing, a guard never evaluated, a set never compared all produce the same clean output as success.

🤖 Generated with Claude Code