Nargo further optimizes cached programs after loading them.

[TomAFrench]

In this snippet we're either compiling the program from scratch or returning a cached artifact. This will then be unconditionally fed into nargo::ops::transform.

noir/tooling/nargo_cli/src/cli/compile_cmd.rs

Lines 182 to 199 in 4ff3081

	let compile_package = |package| {
	let (program, warnings) = compile_program(
	file_manager,
	parsed_files,
	workspace,
	package,
	compile_options,
	load_cached_program(package),
	)?;
	let target_width =
	get_target_width(package.expression_width, compile_options.expression_width);
	let program = nargo::ops::transform_program(program, target_width);
	nargo::ops::check_program(&program)?;
	save_program_to_file(&program.into(), &package.name, workspace.target_directory_path());
	Ok(((), warnings))
	};

This is fine when nargo::ops::transform is a noop on a cached artifact but it seems that MergeOptimizer is not fully optimizing programs on the first instance so when we run the cached program through it again it can further optimize the program resulting in a different bytecode being executed compared to what's saved on disk.