AOT: skip the corpus warmup when the committed prelink cache is current (#24302)

[PurHur]

Fixes #24302 — compiling <?php echo "hi\n"; on a clean checkout took ~9 minutes.

first compile, clean checkout:  517 s  ->  5 s
second compile:                   5 s     (unchanged)

Root cause

warmForUserAotBuild() decided whether to emit the entire helper corpus based on one thing:

$marker = self::coreMarkerPath();     // build/helper-runtime-cache/<prefix><fingerprint>.ok
if (is_file($marker)) { return; }
...
$rc = self::runWarmupCommand($cmd);   // shells out to emit-helper-runtime-object.php

coreMarkerPath() lives under cacheDir() = build/helper-runtime-cache, and that directory is
gitignored (.gitignore:29, 0 files tracked). So a clean checkout can never have the marker, and
the first user AOT build re-emitted the corpus regardless of whether prelinked/helper-runtime/
was present and current.

That is also why #24321 did not move the number: refreshing the committed cache fixes what the
consumption path can use, while the warmup decision never looked at it. Measured 533 s before,
517 s after — ~16 s of ~530.

Found by profiling, not by guessing

My first hypothesis (stale committed fingerprint) was measured and refuted. A SIGALRM sampler on
the real bin/compile.php then collected only 5 ticks in 441 s, top path:

HelperRuntimeCache::runWarmupCommand < warmForUserAotBuild < run < php_compiler_cli_dispatch

The parent is blocked on a subprocess, so parent-side sampling sees almost nothing — the near-empty
profile was itself the signal.

The change

Skip the warmup when the committed per-arch cache exists, carries a core_fingerprint equal to the
live one, and actually contains units. The consumption path already reads prelinkedUnitsDir()
alongside unitsDir(), so in that state there is nothing to emit.

committedCacheIsCurrent() is deliberately conservative — missing manifest, unreadable JSON,
mismatched fingerprint, or an empty units directory all fall through to the warmup. A stale or
partial committed cache is still emitted rather than silently producing a build with missing helpers.

Gate

No CI on lib/, and AOT is informational in compiler-gate.yml, so this is a named local gate.
Branch vs a clean master checkout, run sequentially, both with build/helper-runtime-cache
removed first so the measurement is cold:

	branch	master
cold aot-smoke	8/8	8/8
VM --repeat 2	110/110, exit 0	110/110, exit 0
AOT --repeat 3	16 failing	16 failing

regressions (branch-only): NONE
fixed (master-only)      : NONE

One note on method

My first attempt at that comparison read master's sweep while it was still running (73 of 110
cases) and showed four phantom regressions — j05, j06, k02, m02 — purely because those cases
had not been reached yet. The table above is from the completed run. Worth flagging because a
partially-written results file looks exactly like a real regression set.

What this does not do

It does not make the corpus emit itself faster, and it does not help when the committed cache is
genuinely stale — that path still pays the full ~9 minutes, correctly. Keeping
prelinked/helper-runtime/ current is what makes the fast path available, which is what #24321 and
its release-readiness gate are for.