fix(service): idle awaiting config instead of exit-restart loop when unconfigured

[bigduu]

Summary

Fixes the 1s restart hot-loop reported in #4 when magpie runs as a bamboo-plugin services artifact:

• Config file missing (fresh install, before the operator writes plugin_service_config/magpie/config.json): magpie used to log an ERROR and exit FAILURE. bamboo's ServiceManager restart policy (max_attempts: 0 = unlimited) then respawned it every ~1s forever.
• platforms: [] (config present, nothing enabled): magpie used to warn and exit SUCCESS — but the restart policy restarts clean exits too, producing the same 1s loop.
• Invalid config (parse error / failed validation, e.g. telegram entry missing a token): previously also exited FAILURE and crash-looped.

All three are now treated as a normal "not yet configured" state: magpie stays alive and polls the config path every 30s until it parses, validates, and has at least one platform configured, then proceeds with normal startup.

Change summary

• await_valid_config(path, poll_interval) — the idle loop. Each iteration does one synchronous check (classify_config) then tokio::time::sleep(poll_interval); it never busy-spins. This also preserves the old "accidental hot-reload" the issue calls out: editing the config file while idle is picked up on the next 30s poll, the same way an operator's edit used to land on the restart loop's next spawn.
• classify_config(path) — pure sync classifier (Ready(MagpieConfig) vs Idle(IdleReason)) so the missing-file / parse-error / validation-error / empty-platforms cases are unit-testable without going through the async poll loop.
• Logging: one line when idling begins (or when the failure reason changes), then at most once every 5 minutes while it persists — visible without spamming.
• main() races await_valid_config against the existing shutdown_signal() (tokio::select!), so SIGTERM/Ctrl-C still interrupts a sleeping poll and exits immediately — graceful shutdown is unaffected while idling.
• --check (the interactive GET /api/v1/execute/defaults smoke test) intentionally keeps the old fail-fast behavior: it's an operator-invoked diagnostic, not the supervised service loop, so it should error immediately on a bad config rather than idle.
• Added tokio's test-util feature as a dev-dependency for tokio::time::{pause, advance} in the new tests.

Diff is scoped to src/main.rs (startup/config handling) + Cargo.toml (test-only dependency) — no changes to bridge.rs or the resync/ask-park code (kept separate from the concurrent #9 work in that area).

Test plan

• cargo fmt --check
• cargo clippy --all-targets — clean, no warnings
• cargo build --tests
• cargo test — 172 lib tests + 12 main.rs tests (incl. 6 new: classify_config_* missing/parse/validation/empty-platforms/ready, await_valid_config_does_not_busy_spin_and_picks_up_a_later_config, await_valid_config_resolves_immediately_when_already_valid)
• Manual smoke test of the built binary:
  • missing config file → process stays alive, logs the idle warning, SIGTERM exits cleanly within ~1s
  • platforms: [] → same idle behavior, SIGTERM exits cleanly
• cargo tree -i aws-lc-sys — no match (native-tls pin unaffected by the new dev-dependency)

Closes #4.

https://claude.ai/code/session_014iw5PBsSzDFHus1GfkAK4y