feat(api): idempotent incident creation via dedup_key

[juanjogarciatorres43]

What

Adds server-side de-duplication for incident creation: POST /api/v2/firefighter/incidents now accepts an optional dedup_key, and at most one open incident can exist per key. Automated sources can safely retry/re-deliver without opening duplicate incidents.

This is the IMPACT half of QC-159 (Datadog → IMPACT auto-raise; ADR-0110). It is independently mergeable and fully backward-compatible — manual creation paths (Slack/web) send no dedup_key (NULL) and are unaffected.

Why

The Datadog→IMPACT mechanism re-delivers and the source monitors flap heavily, so without dedup each alert would open a new P3 (and a new Slack channel). Doing dedup at the source of truth (the incident store) is atomic, correct across the incident lifecycle, and reusable by every future adopter — strictly better than a per-receiver dedup store.

How

• Incident.dedup_key — optional CharField (NULL for manual incidents).
• Partial unique constraint — UNIQUE(dedup_key) WHERE dedup_key IS NOT NULL AND _status <= Mitigated (40).
  • Atomic → no duplicate even under concurrent delivery (no TOCTOU).
  • NULL keys never collide.
  • Once an incident reaches Post-mortem/Closed it leaves the partial index, so a later recurrence opens a new incident (no permanent blocking).
• API — serializer accepts write-only dedup_key; on conflict it returns the existing open incident with HTTP 200 (instead of 201) and does not re-trigger the Slack channel workflow.

Tests

• Model: collision while open, Mitigated still blocks, key freed past the open window, NULL never collides.
• Serializer: returns existing incident on conflict / re-raises otherwise.
• ✅ Full Python CI green on the branch (tests on Postgres, mypy, ruff, build).

Notes for review (Pulse)

• Migration 0034 adds a partial unique index on incidents_incident (small table → normal build is fine; switch to CREATE UNIQUE INDEX CONCURRENTLY if a zero-lock build is preferred).
• "Open" window is _status <= 40 (Mitigated) — a recurrence during Post-mortem opens a new incident (intended).
• except IntegrityError is broad but only swallowed when a matching open incident exists; otherwise re-raised.