Allow fallback error handlers to queue commands

[0xEgao]

Objective

Fixes #24821.

Solution

• Pass Commands to ErrorHandler and return it from the handler.
• Apply commands queued by error handlers across executors, observers, combinators, and command error paths.
• Preserve existing scheduling behavior when no error occurs.
• Add a set_error_handler doc example demonstrating how to send AppExit.

[alice-i-cecile]

I really want a feature like this, but this might have surprising performance implications. Can you please benchmark this?

[0xEgao]

I really want a feature like this, but this might have surprising performance implications. Can you please benchmark this?

Yeah sure :)

[0xEgao]

I really want a feature like this, but this might have surprising performance implications. Can you please benchmark this?

Hey sorry for taking time, I benchmarked this against origin/main using Bevy's existing Criterion benchmarks.

Environment:

• Apple M4 MacBook Air (10 cores, 16 GB RAM)
• macOS / Darwin 24.6.0, arm64
• rustc 1.97.0 (2d8144b78 2026-07-07)
• Baseline: origin/main at c7a16115c
• Branch: 9ea9e652e
• Both revisions used the same Cargo.lock and isolated Cargo target directories.

Results:

Benchmark	Mean change (95% CI)	p-value	Criterion result
run_empty_schedule/SingleThreaded	-1.90% [-3.09%, -0.76%]	0.00	Within noise threshold
run_empty_schedule/MultiThreaded	+0.06% [-0.65%, +0.80%]	0.88	No change detected
empty_systems/0_systems	+1.49% [+0.97%, +1.99%]	0.00	Within noise threshold
empty_systems/2_systems	-2.77% [-11.01%, +2.25%]	0.69	No change detected
empty_systems/4_systems	+1.05% [+0.21%, +1.88%]	0.02	Within noise threshold
empty_systems/10_systems	-0.37% [-2.56%, +2.21%]	0.78	No change detected
empty_systems/100_systems	-0.16% [-0.90%, +0.63%]	0.69	No change detected
empty_systems/1000_systems	+0.45% [-2.62%, +3.23%]	0.78	No change detected
param/combinator_system/8_piped_systems	+0.86% [+0.33%, +1.44%]	0.00	Within noise threshold

The first sequential sweep reported regressions for the 10- and 1,000-system cases. I immediately repeated those cases with main and the branch run back-to-back. The regressions did not reproduce: Criterion reported no change (p = 0.78) for both. The table uses those controlled repeat results.

No benchmark was classified as a performance regression by Criterion.

Commands used:

cargo bench --locked -p benches --bench ecs -- run_empty_schedule --save-baseline main
cargo bench --locked -p benches --bench ecs -- run_empty_schedule --baseline main

cargo bench --locked -p benches --bench ecs -- empty_systems --save-baseline main
cargo bench --locked -p benches --bench ecs -- empty_systems --baseline main

cargo bench --locked -p benches --bench ecs -- param/combinator_system/8_piped_systems --save-baseline main
cargo bench --locked -p benches --bench ecs -- param/combinator_system/8_piped_systems --baseline main