fix(z80): legal spill operands and emitted split labels — no invalid asm on fuzz2 (P8)

[oisee]

P8: production Z80 no longer emits invalid assembly on fuzz2-shaped programs

Fixes the root-cause classes behind the assembly failures the differential fuzzer found on main:

• a spill slot used as an ALU or immediate operand;
• spill labels that were referenced but never emitted;
• split labels that were never emitted;
• wide (u24/u32) spill operands;
• an accumulator on the right-hand side of an ALU op;
• SBC through an index register;
• conditional-call intrinsics.

Each fix routes the value through a legal register path. Wide staging and wide slot sizes now apply only to u24/u32 integers. A constant copied into a spilled block parameter is now stored at full width.

Numbers (base 6b8fc39)

• Fuzzer, seeds 0–1099: assembly failures go from 683 to 0. All 1,100 programs assemble, and the Z80 result matches the MIR2 VM on all of them.
  • 5 seeds (56, 399, 784, 812, 967) disagree with the Python oracle. The oracle is right there; VM and Z80 share an older frontend bug: literal-only expressions are typed u8 in a u16 context.
• Critic's independent fuzz, seeds 5000–5499:
  • main: 298 of 500 fail to assemble, and 61 compile to wrong VM values.
  • branch: 0 fail to assemble, 497 are correct, and the other 3 are the same oracle cases.
• Per-assert matrix: 4,016 checks. Passes go from 3,070 to 3,072. 0 newly failing.
• Corpus: assembly failures go from 65 to 61, with no new failures. These now assemble:
  • nanz/nc.nanz
  • nanz/rotozoomer.nanz
  • nanz/test_irc_minimal.nanz
  • glsl_sphere_demo.minz
• Size: files that assemble on both builds grow from 141,594 to 141,630 bytes (+36).
• Bonus fixes:
  • OpCallCond emitted CALL cc with no argument setup (a silent miscompile).
  • highByte(spill) wrote the low byte twice.

Honesty note

Correctness evidence covers fuzz2-shaped programs only. On the critic's second generator (globals, u8 temporaries), production Z80 is wrong on essentially every program, on main and on this branch alike. Bugs that already existed now show up as wrong values instead of assembly errors. The report lists them:

• the memory-parameter ABI mismatch;
• a byte XOR of truncated u16 values computing 0;
• tail-call argument setup clobbering B;
• literal typing;
• shifts on IX/IY halves;
• u32 spill-to-global stores;
• u24 moves through emitMov32.

Those are the next correctness track.

Gate

• Every regression test was seen failing without its fix.
• Required Go gates: exit 0.
• Cross-family critic: MERGE GO after two rounds.

Report: reports/2026-10-02-P8-Invalid-Asm.md.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Wrwo36SzxKRZYTgpDo7hzo

[github-actions]

PR tests: 🟢 pass

Commit: cb307f2b · Full test log

Check	Result
Build mz, mza, mze, mzv	🟢 pass
Stable Go package tests	🟢 pass
Fibonacci compile + assemble	🟢 pass
Nanz language regressions	🟢 pass
Z3 West of House demo	🟢 pass
Z3 small conformance fixtures	🟢 pass

The external CZECH and full Zork I stories are not part of this PR gate.