QWord ⊕ non-negative literal: result/comparison domain demoted to signed

• Type: bug (silent wrong values — signed print, wrong hi() width, wrong ordering). Track P (shared parser typing) — filed per "board = the record"; fixed in the same night session that found it.
• Status: done
• Opened: 2026-07-14, found in the tint642 burn-down right behind [[bug-pascal-qword-to-double-signed]].

What happened

$f0000000 (> maxint) widens to tyInt64 at parse. The equal-width "signed wins" rule then made:

• q + $f0000000 → tyInt64: bits correct, but writeln printed it negative and hi() keyed a 32-bit width off the demoted type ($F303 instead of $FAFAFAFA).
• q > $f0000000 → SIGNED compare: any q ≥ 2^63 ordered wrong.

FPC's range-based constants join the unsigned domain when non-negative.

Fix (same session)

Two parser sites:

• arithmetic binop typing: a non-negative AN_INT_LIT paired with a tyUInt64 operand types the result via (tyUInt64, tyUInt64);
• relational binop: the literal operand node is RETAGGED tyUInt64 so codegen's TypeCompareUnsigned sees two unsigned operands.

Negative literals keep the signed domain (q + (-1) = q - 1 verified).

Verification

test_qword_literal_binop.pas (Makefile-registered) — all rows byte-identical to FPC 3.2.2 output. tint642 now runs through lo/hi and typecasts; its sole residual is {$Q+} overflow-checked arithmetic (skip entry updated).

Log

• 2026-07-14 — resolved, commit 6590369e.