JIT stabilization

[unxmaal]

Core JIT improvements:

• Three-tier JIT (ALU → Loads → Full) with adaptive probe, block chaining (max 32 instrs),
  tier promotion/demotion
• Speculative execution as safety net — all compiled blocks are load-only and speculative,
  catching Cranelift regalloc2 codegen errors via rollback+demotion
• Full-tier blocks terminate before stores (stores go to interpreter)
• Branch-likely instructions (BEQL/BNEL/BLEZL/BGTZL + REGIMM variants) with conditional
  delay-slot nullification via select
• ADDI/DADDI compiled as ADDIU/DADDIU (overflow trap ignored)
• Cross-session profile persistence — deferred lazy replay, kernel-only save, content hash
  validation, boot-settle detection

Infrastructure:

• jit-diag.sh — diagnostic test harness with COW overlays
• codegen_test.rs — standalone regression test suite (1587 lines)
• trace.rs + tools/diff-trace.py — binary trace + differential debugging
• rules/jit/ — 9 rule files documenting hard-won invariants

Result: ~11% JIT coverage, stable IRIX 6.5 desktop (4dwm), ~2.5× throughput vs
interpreter-only.