refactor(server): reduce duplication in server, tls, pubsub, and main

[kacy]

summary

four focused changes across ember-server, targeting repeated patterns that would need to be updated in sync if the underlying logic ever changes.

server.rs — extract setup_tls_listener

the 8-line TLS listener bind block (bind port, load acceptor, log) was duplicated verbatim in both run and run_concurrent. extracted into async fn setup_tls_listener(tls: Option<(SocketAddr, TlsConfig)>) -> Result<...>. both run functions now call it in one line. any future change (e.g. adding SNI, changing the log message) only needs to happen in one place.

pubsub.rs — byte-slice glob matching

glob_match was converting both pattern and input to Vec<char> on every call. under heavy pub/sub load (many pattern subscribers, high publish rate) this was allocating and GC-ing two vecs per match. since all pattern metacharacters (*, ?, [, \) are ASCII, byte-slice comparison is equivalent and avoids both allocations entirely. the backtracking algorithm is unchanged. added doc comments explaining the algorithm.

tls.rs — extract require_file

three near-identical blocks:

if !path.exists() {
    return Err(TlsError::CertFileNotFound(config.cert_file.clone()));
}

extracted to fn require_file(path, make_err). each call site is now a one-liner. adding a fourth file check in the future is a single line.

main.rs — extract exit_err

arg validation failures all followed the pattern eprintln!(...); std::process::exit(1). extracted to fn exit_err(msg: impl Display) -> !. this makes the ! return explicit (the compiler can verify call sites don't need to produce a value) and reduces the visual noise of exit handling throughout main.

what was tested

• cargo build -p ember-server — clean
• cargo test -p ember-server — 43 tests pass, 0 failures
• cargo fmt --check -p ember-server — no formatting issues
• manual smoke test: server starts, redis-cli PING returns PONG, PSUBSCRIBE + PUBLISH delivers to pattern subscriber

design considerations

the setup_tls_listener extraction handles the simplest clearly-duplicated code. the accept loop itself differs between run and run_concurrent in the spawned task body (different handler function, different captured state) — extracting that would require boxing futures or trait objects, which adds complexity. left as a follow-up if warranted.