fix(mk): a warned assembly must not fail the build

[mgrossmann]

as370 returns IFOX00's severity as its exit status — 0 clean, 4 warning, 8
error, 12 severe, 16 terminal — and a warned assembly still punches its deck. On
MVS that was COND=(8,LT): the ASM step returned 4 and the linkage editor ran
anyway.

make stops on any non-zero status, so one warned card would fail a whole build
over a diagnostic the assembler itself considers survivable.

$(Q)$(AS) $(ASFLAGS) -o $@ $< || { rc=$$?; [ $$rc -lt 8 ] || exit $$rc; }

Verified: clean (RC 0) builds, warned (RC 4) builds with the diagnostic on
stderr, errored (RC 8) fails as make: *** [build/bad.o] Error 8 — the true
severity survives into make's message.

Follow-ups, after this merged

• fix(as370): a discarded statement is an error, not a severity-4 warning cc370#84 is the necessary companion and is still open. The
  nsf370 team showed the hole this rule has on its own: IFO026 is severity 4,
  so RC < 8 swallows it — and IFO026 is not an ordinary warning, it means a
  card was eaten and the deck was punched anyway. Sent through exactly this rule,
  a re-lengthened comment card that eats a DCBD gives exit code: 0 and three
  corrupted bytes. cc370#84 splits the two cases in the assembler, where the
  distinction exists: a consumed comment stays severity 4, a consumed
  statement becomes 8. Until it lands, this rule tolerates both.
• .DELETE_ON_ERROR: followed in c80e23c: as370 writes its deck even when it
  flags the assembly, so a build stopped at RC 8 was leaving a .o newer than
  its source, which the next make called up to date and linked.

Not addressed here

The .c rule cannot do the same: the cc370 driver treats any non-zero status
from the assembler it invokes as fatal, and that is inside the driver. Nothing
hits it today — generated assembler carries no column-72 comment cards.

[mgrossmann]

Hold this until mvslovers/cc370#84 lands, and merge it after.

The nsf370 team showed the hole in what this PR does on its own: IFO026 is
severity 4, so RC < 8 swallows it — and IFO026 is not an ordinary warning,
it means a card was eaten and the deck was punched anyway. Sent through
exactly the rule below, a re-lengthened comment card that eats a DCBD gives
exit code: 0 and three corrupted bytes. That is the detection this PR would
have removed.

cc370#84 splits the two cases in the assembler, where the distinction actually
exists: a consumed comment card stays IFO026 at severity 4, a consumed
statement becomes severity 8. With that in place, RC < 8 here tolerates
only what is genuinely harmless — and every project gets the protection, not
just the ones that add a card-column lint.

Measured after the split: libc370 743 modules clean, rexx370 three warnings and
no new error, nsf370's two broken modules RC 8 (their fixed branch clean).