Verifying a decision routine by replaying its calls

What to know

2026-10-24, Hardwar's pilots_gen.py, crates/pilots/tests/decide.rs.

PilGenControls (0x466c64) is 2,600 instructions of decisions about one pilot after another, and nearly every decision asks another system something: the vehicle, the trade board, the clock, the garrison. Building a real world for it in the emulator would mean porting half the game first. Three practices made it testable bit for bit without that.

Stub every callee and record the call, not just the answer. Each routine the step calls is a stub that draws a random answer, writes random results through its pointer arguments, and appends a record: the address, EAX, EDX, EBX, ECX, the stack arguments, the answer and the results. The port calls the same routines through a trait, one method per original routine; the test's implementation of the trait takes the next record, checks that the address and the arguments that carry values (not frame pointers) are the port's, and hands back the original's answer. A wrong branch shows up as the first call that differs, with its index. It also makes the order of calls, and so the order of every random draw made inside them, part of what is verified.

Keep pointers as pointers. The earlier pilot tests wrote pointers as small indices. Here a plan's word 1 is sometimes a 16-bit good written over a building pointer left in the buffer, so its high half is half a pointer; an index encoding turns that into a word the port cannot reproduce. Writing the emulator's addresses as they are, and giving the port the same handles, fixed it.

Model the frame when the routine reuses it. The routine builds plans in a handful of stack buffers and sets only some words of each; the rest is whatever the previous pilot left. The port keeps those buffers (decide::Frame) and the emulator zeroes the stack before each case, so only the frame's state before the first pilot is a deviation.

A pitfall on the way: filtering a long listing to read it faster (dropping inc eax, mov edx, ebp and the like as loop noise) also dropped a real inc eax, and the port first left out a day carry that the original has. The emulator found it on the 11th case; read the unfiltered listing for anything the port depends on.