feat(sinks): Slack/Discord webhook sink for state-transition notifications

[jiunbae]

Summary

• New [sinks.webhook] HTTP sink that POSTs to a Slack/Discord webhook on agent state transitions, so operators get phone alerts the moment an agent flips to WaitingInput/Error while they're AFK from muxa watch.
• Slack ({"text": "..."}) and Discord ({"content": "..."}) wire shapes auto-detected from the URL; flavor = "generic" posts the full Transition JSON for n8n/Zapier-style receivers.
• Best-effort: failed POSTs log at WARN and drop. No on-disk queue, no retry backoff.

Config

[sinks.webhook]
enabled         = true
endpoint        = "https://hooks.slack.com/services/T0/B0/abc"   # OR
endpoint_env    = "MUXA_SLACK_URL"                               # preferred
flavor          = "slack"            # slack | discord | generic
on_states       = ["WaitingInput", "Error"]
rate_limit_secs = 60

endpoint_env wins when both are set — the webhook URL is the secret on Slack/Discord, so the env-var path keeps it out of shared dotfiles.

Filter + rate-limit

Two intentional pieces of policy live in this sink:

1. on_states filter (default ["WaitingInput", "Error"]) — most transitions are routine Idle ↔ Working and would spam.
2. Per-(kind, session_id, state) rate-limit (default 60s) — kept in an in-task HashMap. WaitingInput ↔ Working flap-flap is common during permission-grant loops; without the limiter one flaky agent fires 30 push notifications a minute.

We don't queue or retry on failure: Slack/Discord both retry on their side, and a buffered alert delivered ten minutes late is worse than no alert at all (operators already moved on). One bounded HashMap per task, no shared state.

Test plan

• cargo build --workspace
• cargo test --workspace (260 + 7 + 177 + 8 + 2 unit/e2e tests pass)
• cargo clippy --workspace --all-targets -- -D warnings
• cargo fmt --all -- --check
• New unit tests cover: format_message_for_waiting_input, format_message_for_error, should_forward_filters_idle_transitions, rate-limiter suppress/release/per-key behaviour, flavor_inference_from_url, payload shape per flavor (Slack text / Discord content / Generic full Transition), config validation (enabled requires endpoint or endpoint_env, env wins over inline, invalid URL rejected).
• Manual: enable with a real Slack incoming-webhook URL, run muxa watch, trigger a WaitingInput → confirm push arrives once, then again on the next WaitingInput outside the rate-limit window.
• Manual: same with a Discord webhook URL — verify the body uses content, not text.

🤖 Generated with Claude Code