Sweep: let a case declare a backend it cannot pass (#23779), and correct my false g07 claim

[PurHur]

Why

The AOT differential sweep cannot function as a gate while cases that exercise unimplemented
features sit in it failing forever. Three such cases are mine, from #23754.

The mechanism

An opt-in marker with a mandatory reason:

// @differential-skip-aot: print_r() JIT helper requires Runtime->vm from thin standalone init (#9190 / #23540)

The reason prints with the skip and the summary reports the skipped count, so a skip can never be
silent:

skip    g02_print_r.php    print_r() JIT helper requires Runtime->vm from thin standalone init (#9190 / #23540)

29/52 match Zend (aot backend, 3 skipped)

AGENTS.md documents it with the warning attached — only for genuinely unsupported features, never
to silence a real defect, because a skip marker is the cheap green of §4 wearing a different hat.
It also now records that --aot is not green (27/53 when first run end-to-end, #23779), so its
raw count is not a pass/fail signal.

Cases marked

case	reason
g01_var_dump	var_dump() of string/bool/null needs Runtime->vm (#23540)
g02_print_r	print_r() needs Runtime->vm (#9190 / #23540)
e29_var_dump_indent	var_dump() of arrays needs Runtime->vm (#23540)
g07_incdec_resource_provenance	different — a real defect, #23811

The first three are unsupported features. g07 is not: it fails AOT because of a genuine bug,
and the marker says so explicitly and tells you not to read it as "this shape works under AOT". It
is skipped rather than deleted so that removing the marker restores an AOT guard the moment #23811
lands.

Correcting a false claim of mine

#23781 states g07 "passes VM and AOT". That was wrong. I created the file while the sweep I
cited was already running, so its case list had been globbed before the file existed — g07
appears in neither section of that log. I read "absent from the failure list" as "passed".

Re-measured with the case genuinely included: VM passes, AOT fails. And the AOT failure is not
caused by #23781 — with lib/ reverted to 544d1dca9 the same program dies rc=134/SIGABRT with
no output at all, versus wrong-but-partial output after. The fix un-masked pre-existing breakage.
Filed as #23811 (integer rendered as Resource id #N when its value collides with a live handle,
plus a mis-evaluated decrement). Full correction posted on #23781.

Gates

• VM: 55/55 match Zend, exit 0 — and zero skips fired, since the marker is backend-specific,
  so all four cases still gate the VM path exactly as before.
• AOT: 29/52 match, 3 skipped, exit 23. Failing case names compared against the pre-AOT: emit concat for unallocated call-arg Temporaries (#23779) #23792
  list: newly passing are c04_concat, d04_concat_dim, e09_nested_calls (fixed by AOT: emit concat for unallocated call-arg Temporaries (#23779) #23792 /
  AOT: fix min()/max() boxed variadic compare (#23779) #23801); nothing newly fails except the two g07* cases, which were never in the earlier run.

No lib/ changes.