I found it. I had to throw some continues_on() else it'd forget where it started and go on the default serial scheduler.
Wrote a summary of the experiment if you haven't seen it: mropert.github.io/2025/11/21/t...

I was in fact. And possibly using it wrong, but tbh it's really hard to tell if I'm constructing the job graph in a way that hinder the parallelism I intented.
What's your recommended way to debug those kinds of issues?

So far observations are: if I have a some prior and follow up tasks it goes serial.
If I just do single bulk() it does it in parallel.
If I do a wait_all() on 2 bulks it does one parallel, then the other (not both simultaneously).
And debugging _why_ it doesn't parallelize as expected is hard.

Do not use an app for anything that doesn't strictly needs an app.
This is free professional advice from a senior software engineering consultant.

Do intro mission.
Get broken leg, limp forever until I find splint.
Look for the ledx everyone mentions.
You don't keep your inventory anymore.
Talk to Ragman.
He looks like he's renderered on a PS2.
Talk to Therapist.
Game bugs out I can't exit or get dialog option.
Alt+F4.

Yes I have validation layers on and no they don't seem to ever catch anything.

You can make it UB in release mode for performance that's fine, but for god's sake add some asserts or contract checks in debug.

On literally the same day 🤡
www.gamesindustry.biz/sharkmob-con...

On literally the same day 🤡
www.gamesindustry.biz/sharkmob-con...