Adom Hydrogen

Likely cause and a fix on main (bbb77117): the main thread's stack. Windows gives it 1 MB, and in a debug build the async state machines and Tauri's setup closures are many times larger than in release, which fits a debug exe dying with 0xC00000FD shortly after window creation while the release build of the same tree runs.

hydrogen-app/build.rs now links debug builds on MSVC with /STACK:8388608 (8 MB reserved; pages are committed only as used). The release profile is untouched, so shipped installers are byte-for-byte the same as before this change.

I could not reproduce it here (AdomLapper runs debug builds without the crash), so could you rebuild the debug exe on Colby-ROG2 from main and say whether it stays up? If it still dies, the frame from the dump (the PDB is in target/debug) tells us which state to box instead.

Tested on Colby-ROG2 from main bbb77117. The /STACK:8388608 is in the exe (PE SizeOfStackReserve 8388608), and the debug build still dies the same way: 0xC00000FD about half a second after window creation.

The dump says why: the overflow is not on the main thread. cdb (WinDbg 1.2606, symbols from target\debug) shows thread 42, "tokio-rt-worker", faulting in __chkstk on entry to hydrogen_control::handle_control_inner::async_fn$0::async_block$1, under handle_control -> hyper http1 dispatch -> hydrogen_control_ai_threads::ai_tabs::start_control_server. The stack is only about 60 frames deep, so this is not recursion: it is the router's single poll frame, which in a debug build (no stack-slot reuse across the match arms) is bigger than a tokio worker's 2 MB stack. It fires on the first control-API request after the window opens.

Confirming experiment: the same debug exe launched with RUST_MIN_STACK=16777216 in its environment stays up (113 s and answering /buildinfo, dev=true). So tokio's workers take std's default stack, which that variable controls.

Two ways to fix it, your call:

1. Debug only, smallest: in main(), before the Tauri builder starts any thread, #[cfg(debug_assertions)] set RUST_MIN_STACK to 16 MB when it is unset (std reads it once, at the first spawn). Release is untouched, and the /STACK change can come back out since the main thread was never the one overflowing.
2. Root cause: keep the router's per-arm work out of handle_control_inner's own frame (move the heavy arms into separate non-inlined async fns). That shrinks the frame in release too, where it is smaller but still large.

Why AdomLapper does not crash is worth one check: whether RUST_MIN_STACK is set in your environment there, or your toolchain differs (this box is rustc 1.95.0).

Great dump work, thank you: my main-thread guess was wrong, and option 1 is on main as 91d08849.

• main() now sets RUST_MIN_STACK=16777216 under #[cfg(debug_assertions)], as its first statement and only when the variable is unset, so it precedes every thread spawn and an explicit value in the environment still wins. Release builds never compile it.
• The /STACK link arg from bbb77117 is gone from build.rs, since the main thread was never the one overflowing.

Why AdomLapper does not crash: RUST_MIN_STACK is unset there (user and machine), and its toolchain is rustc 1.93.1 against your 1.95.0, so the likeliest difference is codegen: 1.95 lays out that debug poll frame larger. I am leaving option 2 (moving the router's heavy arms into their own non-inlined async fns) for a separate change, because it touches every control route and release is not overflowing today.

Could you rebuild the debug exe from main on Colby-ROG2 and confirm it stays up with no RUST_MIN_STACK in the environment? I will close this on your confirmation.

Confirmed on Colby-ROG2. Debug build of main 91d08849 (rebuilt, not a relink), RUST_MIN_STACK unset in the user and machine environment: hd_launch reported stillRunning after its 30 s settle, and at 60 s the process was up and /buildinfo answered 91d08849 dev=true. Before the fix the same box died with 0xC00000FD half a second after the window opened. Good to close.

Thanks Colby. Confirmed fixed by 91d08849 (debug builds set RUST_MIN_STACK to 16 MB before any thread starts; release unchanged). Closing.