# Step 11.1 — real-comparison codegen fix (tag: `step11.1`)

Real `< > <= >=` never emitted `D5` (`0D5H`) — only `=` / `#`
did. `RealLt/Le/Gt/Ge` (comments even said `[gt lt] -> ...`,
inputs no caller provides) emitted bare stack juggling, leaving
a raw operand bit-pattern as the boolean, so e.g. `-1.0 < 0.0`
read back FALSE. The same recipe was shared with m2compiler-V2
(found via its step-5.2 probe; full story in its
`docs/summary_m2comp_step5.2.md`). `D5` itself was always
correct — no VM change.

Fix: one `OPrCmp` prepended per procedure (contracts now
`[r1 r2]`). New regression `tests/t_rcmp.mod` (13 cases incl.
negative-vs-zero and `-0.0`, ExitCode 31). Suite 158/158 with no
other expectation touched (`t_real` passes genuinely now).