…+ email

Closes the operator-feedback loop. The cockpit-complete observation
from #145: every panel now reads real data, but Joshua still has
to LOOK at the cockpit to see alerts. This bucket flips that — when
a critical/warning alert fires, Kora pings Joshua via Slack DM
(immediate); info alerts go to email.

New module: ``kora_cli/alerts/notifier.py``

  * ``AlertNotifier`` class — periodic-task entry point
    ``run_notification_cycle`` calls #145's aggregator, computes
    set-diff against last cycle's alert IDs, dispatches only
    newly-firing alerts.
  * Channel routing (PM §4 Q2 default): critical/warning → Slack
    DM, info → email.
  * Dedup state in-memory only (PM §4 Q3 default): empty on first
    cycle → all currently-active alerts fire as "new" (one-time
    re-ping on restart; persistence is a future bucket).
  * Failed dispatch still adds alert ID to last_alert_ids (no
    spam on transient SMTP/Slack errors).
  * Audit emit ``notification.dispatched`` per dispatch attempt
    (success OR failure) — operator triages via audit panel if
    expecting a notification that didn't arrive.
  * Pure formatting helpers (Slack DM text, email subject, email
    body, relative-time) extracted for isolated unit-testing.

New listener: ``kora_cli/listeners/alert_notifier_listener.py``

  * Constructs the notifier bound to ``current_slack_client`` +
    ``current_purelymail_client`` lazy factories (matches the
    accessor pattern from KR-MCP-SEND-TOOLS).
  * Periodic task ``alerts.notify`` registered at import time @
    180s default (PM §4 Q1); ``KORA_ALERT_NOTIFY_INTERVAL_SEC``
    env override.
  * Shutdown resets dedup state so a subsequent listener start
    sees a clean slate.
  * Defense-in-depth outer try/except in run_notification_cycle
    so any path that bypasses the notifier's inner catch can't
    crash the heartbeat scheduler.

Audit seam extension: ``notification.dispatched`` added to the
SeamName Literal in ``kora_cli/audit/jsonl_sink.py`` (small +
additive; existing seams unchanged).

K-DG verification at HEAD ``2345d51`` (matches the spec's cited
SHA): all 5 accessors confirmed in place + interfaces matched
(compute_active_alerts ✓ in #145 / current_slack_client /
current_purelymail_client / register_periodic_task ✓ in heartbeat
scheduler / emit_audit ✓ in audit/jsonl_sink).

50 new tests pass:
  * 36 notifier tests (formatting + routing + dedup + failure
    isolation + audit shape + telemetry + dispatch outcome)
  * 14 listener tests (registration + cadence resolution +
    lifecycle + run_notification_cycle short-circuits +
    defense-in-depth + factory wiring)

Cross-bucket regression: 600/600 when run serially. With
pytest-xdist parallelism, 5-ish flaky failures appear in
test_email_inbound_handler.py — VERIFIED pre-existing on bare
HEAD without these changes (3 runs of bare HEAD reproduced 4/6/0
failures, same email_inbound tests). Not introduced by this PR;
filed as a separate xdist-ordering pollution issue for the
follow-on bucket queue.

Ruff clean.

§4 PM-open status — all DEFAULTS applied in ST1:
  Q1 cadence 180s (3 min)               — accepted
  Q2 routing critical/warning→Slack, info→email — accepted
  Q3 fire on first cycle (no persistence) — accepted
  Q4 kora__send_test_alert MCP tool      — DEFERRED to ST2

After ST2 lands (per-category cooldown + burst dampening + digest
mode + operator runbook + kora__send_test_alert MCP tool), the
operator-feedback loop closes fully.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>