fix(as370): a discarded statement is an error, not a severity-4 warning

[mgrossmann]

The nsf370 team found the hole, and they found it in the rule I proposed for mbt
(mvslovers/mbt#94, accept RC < 8). Re-lengthening the comment card that had
eaten their DCBD and sending it through exactly that rule:

make would see exit code: 0        <-- the build PASSES
CORRUPTED: 3 bytes differ from the intended code

The defect that turned their CI red at noon would have passed silently by
evening. They are right, and the fix belongs here rather than in a per-project
lint.

What IFOX00 conflates

Severity 4 covers two different things:

	costs
a continued comment eats another comment	nothing
a continued card eats a statement	the statement — and the deck is punched anyway

Same message, same severity. On MVS that was survivable: COND=(8,LT) let RC 4
through and a test run would find the damage. In a host build, where the exit
status is the only channel, it is not.

The split

as370 keeps IFOX00's number, message and severity for the harmless case and
raises the one that loses a statement to 8. The bytes stay IFOX00's — only
the return code says a build must stop. There is precedent in the file: a
construct that is valid Assembler XF but unimplemented here already returns 8
with the note "not IFOX's error, but it must not pass silently".

The test is the consumed card, not the statement that ate it:

• a comment card is never a loss — as a comment it generates nothing, and
  merged into an operand it only spoils a TITLE string. rexx370's
  trxldc.asm and tlnkterm.asm are that shape (a TITLE whose quoted text
  runs past column 71 eats the comment below it) and stay warnings. This is the
  case my first cut got wrong;
• a card eaten by a continued comment is discarded whole → error, including
  when IFOX00 does not warn at all, which is what happens when the card's
  columns 1-15 are blank — a macro operand card vanishing without a word;
• a card eaten by a continued statement keeps only columns 16-71, so a label
  or operation in 1-15 is thrown away → error. That is rs_goodidx.s losing its
  own END.

Measured

project	RC 4	RC 8	decks
libc370 (743)	0	0	none moved
rexx370 (18)	3	1 — irxprobe.asm's pre-existing AMODE/RMODE gap, zero cards lost	—
nsf370 main (9)	7	2 — exactly the modules that lost a DCBD, two EQUs and an instruction	—
nsf370 PR branch (9)	0	0	—

tests/cont72.s now returns 8 where IFOX00 returns 4, deliberately and with the
reason in the fixture: its two eaten statements are errors, its three
comment-eats-comment cards and its IFO069 stay warnings, and the section is
still the oracle's eight bytes. The listing reference is unchanged and still
carries IFOX00's own verdict.

Order with mvslovers/mbt#94

This should land first. With the split in place, RC < 8 in the mbt rule
tolerates only what is genuinely harmless, and a build that drops a statement
still stops — for every project, without each one having to add a lint of its
own. The nsf370 team's check-card-columns.sh remains worth having: it catches
the card before the assembler runs, and it covers sources the assembler never
sees.