So suppose I wanted to write a pd-very-lite with no audio, just messages. The only objects would be 'bang', 'delay' and some kind of output (e.g. sending a midi CC or something). What sort of code do you have to write to get a reliable ms delay of an event?

More generally, again pedagogical value for now, how does one sequence midi events, or events in general, either when there is an audio thread to sync to, or without.

clock_delay will take the pointer to t_clock and put it on the scheduling list, which is run through every 64 samples to the next 64-sample time

when you create the clock, with clock_new, you give it a callback "tick" function that will be called when the scheduler reaches that point in the scheduling list.

here are the time-based objects: https://github.com/pure-data/pure-data/blob/master/src/x_time.c

your scheduler basically has to tick every so often (and either sleep or spin in between), going through the scheduling queue (in pd this is pd_clock_setlist in m_sched.c) and popping off of the queue for every time value less than the next tick. In between your ticks get input from the ui and schedule things on the queue based on that.

I forget the details, but when audio is running the 'ticks' are based on calling the audio output api (or visa-versa?) every 64 samples.
you can see pd's scheduler here: https://github.com/pure-data/pure-data/blob/master/src/m_sched.c

getting an accurate sleep can also be tricky, there are some articles on it.

short answer: you have to make sure that the data runs in the high priority thread.

pd and max have a main thread, a high priority thread (overdrive), and an audio thread (among others)

how to do that in C++ is surely documented in the SDK examples, but i will not be of much help here.

maybe compare the source code of [+ ] vs. [delay], the latter outputs to the high priority thread if one exists.

(scheduling events works fine: metro 1 -> click~ at vectorsizes 32 does what it is supposed to - in max it required the hp-thread to be available or it would produce hiccup.)

flag -noaudio

@Obineg all of pd vanilla runs on 1 thread (well, the 'core' non-wish process anyways)
Message processing is simply interspersed w/ audio processing.

Almost. Pd does use a single thread for DSP and message processing, but this is not the same as the audio callback (unless "use callbacks" is enabled), so you have (at least) two threads.

Side note: In Pd, the audio callback does almost no work; it just reads/writes samples from/to a ringbuffer and returns. Pd's main thread waits on the ringbuffer (see sys_send_dacs), so it is indirectly (asynchronously) driven by the audio callback.