feat(private-server): agent guidance + published flag for MCP incidents

[passcod]

🤖 Gives MCP agents the context to interpret incidents correctly — they were over-counting raw flapping rows as real incidents.

Stacked on #277 (base will retarget to main once that merges).

What an agent gets

• Each incident now carries published — true when it actually surfaced to operators (its Slack open notice was delivered), which happens only when it outlived its group's grace window (slack_open_delay, ~3 min default) or it escalated (a critical issue joined, bypassing the grace). find_incidents also returns published_count, and each incident gets open_duration_secs.
• published is derived from the slack_outbox (new SlackOutbox::delivered_open_ids) — the authoritative "it surfaced" signal. That's exact: it accounts for opens cancelled when an incident flaps shut within the grace, and for delivery give-ups. It is NOT a now - opened_at >= 10min approximation (the grace is per-group and defaults to 3 min, and a time guess would misclassify cancelled/failed deliveries).

The help text

• The find_incidents tool description and the server-level instructions (returned at initialize) now explain that find_incidents includes heavy sub-grace flapping, that event_count is raw churn (not duration or severity — a huge count can be a sub-minute flap), and that summaries/rankings should count published incidents, not raw rows.

This is the mechanism for "give agents context": MCP clients read the server instructions on connect and the per-tool descriptions, so the guidance travels with the tools.

Spec updated; MCP tests extended to cover published / published_count.