We had a rather scattered selection of delay functions being used:

• delay simply took a number of loop iterations, for a C loop with a nop in it. Analysis of this loop as compiled by GCC 15.2 shows it currently takes 7 cycles per iteration on the M4. This function has been removed, and all uses of it have been replaced with calls to delay_us or delay_ms according to either apparent intent, or actual effect, at each call site.

00004618 <delay>:
    4618:       2300            movs    r3, #0
    461a:       4283            cmp     r3, r0           // 1 cycle
    461c:       d100            bne.n   4620 <delay+0x8> // 2 cycles, branch taken
    461e:       4770            bx      lr               // skipped
    4620:       bf00            nop                      // 1 cycle
    4622:       3301            adds    r3, #1           // 1 cycle
    4624:       e7f9            b.n     461a <delay+0x2> // 2 cycles, branch taken

• delayms in rad1o/display.c, which builds on delay but "naively assumes 204 MHz instruction cycle clock and five instructions per count" (wrong, though might have been right on earlier compilers). Replaced with delay_ms.
• portapack_sleep_milliseconds, which did the same. Replaced with delay_ms.
• delay_us_at_mhz, which I added to provide an actually-calibrated loop, but which requires knowing the current CPU clock speed. Replaced with delay_us, which uses delay_us_at_mhz internally, but with the clock speed parameter provided by the new cpu_clock_mhz symbol, updated by code in cpu_clock when making changes.
• halt_and_flash passed its argument to delay. Replaced that call with delay_ms, and updated arguments to halt_and_flash accordingly.