perf(core): flatten demux filter chain into a vector

[Leonqn]

Summary

• Replace nested closure chain in Demux with a flat SmallVec of filters
• Add mailbox benchmark measuring per-call cost of send/try_send (routed and direct) across 1–100 consumer groups
• Enable full feature for bench crate (needed for elfo-configurer)

Previously each route_to wrapped the previous filter in a new closure, creating a recursive call chain. With many groups (50+) this caused overhead due to deep indirect calls and CPU cache misses. The flat vector iterates filters sequentially.

Benchmark results

send_routed (ns per call, system allocator):

Groups	Nested closures	Flat vector	Change
1	227	230	~0%
10	735	716	~0%
30	812	822	~0%
50	1106	847	-24%
75	1274	874	-32%
100	1670	883	-47%

try_send_routed (ns per call, system allocator):

Groups	Nested closures	Flat vector	Change
1	182	179	~0%
10	717	692	~0%
30	794	759	-5%
50	1175	777	-34%
75	1351	809	-40%
100	1694	841	-50%

Test plan

• All existing tests pass
• New mailbox_sys benchmark passes (cargo bench --bench mailbox_sys -- --test)
• Existing messaging_sys benchmark compiles and runs

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[loyd]

I've also tried using the dynstack crate instead to avoid Arc<dyn Fn>, but it seems the difference is negligible:

group                      b188                                   dynstack
-----                      ----                                   --------
send_routed/100g32w        1.01   612.5±66.64ns        ? ?/sec    1.00  607.9±113.70ns        ? ?/sec
send_routed/10g32w         1.04   507.8±92.74ns        ? ?/sec    1.00   489.9±69.34ns        ? ?/sec
send_routed/1g32w          1.04   173.8±68.22ns        ? ?/sec    1.00   167.1±48.85ns        ? ?/sec
send_routed/20g32w         1.01   517.3±60.31ns        ? ?/sec    1.00   514.1±80.83ns        ? ?/sec
send_routed/2g32w          1.00   329.1±47.56ns        ? ?/sec    1.05   344.9±95.70ns        ? ?/sec
send_routed/30g32w         1.03   532.9±92.16ns        ? ?/sec    1.00   519.2±80.79ns        ? ?/sec
send_routed/50g32w         1.05  599.9±152.40ns        ? ?/sec    1.00  569.4±117.64ns        ? ?/sec
send_routed/5g32w          1.00   426.5±45.50ns        ? ?/sec    1.05   446.3±66.93ns        ? ?/sec
send_routed/75g32w         1.00  636.7±151.27ns        ? ?/sec    1.02  652.2±300.52ns        ? ?/sec
try_send_routed/100g32w    1.06  652.3±187.36ns        ? ?/sec    1.00  613.2±172.57ns        ? ?/sec
try_send_routed/10g32w     1.00   499.8±93.48ns        ? ?/sec    1.02   511.0±96.79ns        ? ?/sec
try_send_routed/1g32w      1.07   146.7±39.16ns        ? ?/sec    1.00   137.2±19.96ns        ? ?/sec
try_send_routed/20g32w     1.02   521.3±84.37ns        ? ?/sec    1.00   512.6±61.14ns        ? ?/sec
try_send_routed/2g32w      1.05   336.1±66.64ns        ? ?/sec    1.00   320.4±79.08ns        ? ?/sec
try_send_routed/30g32w     1.00   522.4±54.15ns        ? ?/sec    1.00   524.5±69.52ns        ? ?/sec
try_send_routed/50g32w     1.08  593.7±222.96ns        ? ?/sec    1.00   551.6±77.44ns        ? ?/sec
try_send_routed/5g32w      1.08   464.5±53.77ns        ? ?/sec    1.00   430.8±62.59ns        ? ?/sec
try_send_routed/75g32w     1.00   583.7±84.81ns        ? ?/sec    1.02  593.4±202.98ns        ? ?/sec

[loyd]

I decided to merge these changes without the mailbox.rs benchmark because it mostly duplicates the existing messaging.rs, adding only not very common axes like the number of groups.