feat(sfm): per-unit event history with attribution + Unattributed bucket

Phase 2 of the SFM integration.  Adds a "SFM Events" section to the
seismograph unit detail page (/unit/{id}).  Every event SFM has for the
serial is shown, with each event annotated by which project/location
assignment window it falls into.  Events outside every assignment window
get the "⚠ Unattributed" badge plus a "<N>d before/after <nearest location>"
hint — that's the operator's signal that backdating an assignment (Phase 1
edit-pencil) will absorb the orphan events.

Backend:
- backend/services/sfm_events.py: new events_for_unit() helper.  Fetches
  all events for the serial via SFM /db/events (one call, ceiling 5000),
  loads every UnitAssignment for the unit + resolves MonitoringLocation +
  Project names, then annotates each event with attribution or
  nearest_assignment (signed delta_days).  Bucket filter: all /
  attributed / unattributed.  Stats always reflect the full event set so
  the "Unattributed" KPI tile is meaningful regardless of which bucket
  is being viewed.

- backend/routers/units.py: new GET /api/units/{unit_id}/events with
  bucket / date-range / false_trigger / limit query params.  404s on
  unknown unit_id; returns an empty payload for non-seismograph
  device_types so the page can render the section conditionally.

Frontend (templates/unit_detail.html):
- New "SFM Events" section between "Deployment History" and "Timeline",
  styled to match the existing card pattern (border-t divider, same
  heading weight).
- Hidden by default; revealed only when currentUnit.device_type ===
  'seismograph' after the unit data loads.
- Four KPI tiles: Total Events / Unattributed (highlighted amber when
  > 0) / Peak PVS / Last Event.
- Filters: Bucket (all|attributed|unattributed), From/To, False
  Triggers, Limit, + Refresh.
- Event table with Attribution column.  Attributed rows link to the
  project/location detail page; unattributed rows are tinted amber
  and show "<N>d before/after <nearest location>" with a link to the
  nearest location.
- Empty-state copy varies by bucket: e.g. unattributed-with-zero shows
  "✅ All events for this unit are attributed to a project/location".

Verified end-to-end against BE11529 (81 events total, 24 attributed,
57 unattributed — all 57 unattributed events emitted within hours of
the assignment start, which means backdating the assignment by a day
would attribute every one of them).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/services/sfm_events.py

@@ -29,7 +29,7 @@ from typing import Optional
 import httpx
 from sqlalchemy.orm import Session
 
-from backend.models import UnitAssignment, RosterUnit
+from backend.models import UnitAssignment, RosterUnit, MonitoringLocation, Project
 
 log = logging.getLogger("backend.services.sfm_events")
 
@@ -252,6 +252,168 @@ async def events_for_location(
     }
 
 
+# ── Per-unit (cross-project) view ─────────────────────────────────────────────
+
+
+async def events_for_unit(
+    db: Session,
+    unit_id: str,
+    *,
+    bucket: str = "all",   # "all" | "attributed" | "unattributed"
+    from_dt: Optional[datetime] = None,
+    to_dt: Optional[datetime] = None,
+    false_trigger: Optional[bool] = None,
+    limit: int = 500,
+) -> dict:
+    """Return events for a unit annotated with their assignment attribution.
+
+    Unlike events_for_location (which queries SFM per assignment window), this
+    helper queries SFM for ALL events for the serial within the optional
+    [from_dt, to_dt] filter, then walks each event against the unit's
+    UnitAssignment intervals to compute attribution.
+
+    Bucket semantics:
+      - "all":          every event, attributed or not
+      - "attributed":   events that fall inside at least one assignment window
+      - "unattributed": events with no overlapping assignment (the diagnostic
+                        bucket — operator should fix assignment dates to
+                        attribute these)
+
+    Each event gets an extra `attribution` field:
+      {assignment_id, location_id, location_name, project_id, project_name,
+       assigned_at, assigned_until}                                 or None
+
+    Unattributed events also get a `nearest_assignment` field with the
+    same shape plus `delta_days` (signed; negative = event before assignment).
+    """
+    # 1. Pull all assignments for this unit (any device_type — caller has
+    #    already filtered by seismograph in the route).  Order matters: we
+    #    want the earliest-start assignment first so attribution prefers the
+    #    chronologically-first overlap when there are simultaneous active
+    #    assignments at different locations (rare but possible).
+    assignments = (
+        db.query(UnitAssignment)
+        .filter(UnitAssignment.unit_id == unit_id)
+        .order_by(UnitAssignment.assigned_at.asc())
+        .all()
+    )
+
+    # Resolve location + project names once.
+    loc_ids = {a.location_id for a in assignments}
+    proj_ids = {a.project_id for a in assignments}
+    loc_map = {
+        l.id: l for l in db.query(MonitoringLocation).filter(
+            MonitoringLocation.id.in_(loc_ids)
+        ).all()
+    } if loc_ids else {}
+    proj_map = {
+        p.id: p for p in db.query(Project).filter(
+            Project.id.in_(proj_ids)
+        ).all()
+    } if proj_ids else {}
+
+    now = datetime.utcnow()
+
+    def _attr_dict(a: UnitAssignment) -> dict:
+        loc = loc_map.get(a.location_id)
+        proj = proj_map.get(a.project_id)
+        return {
+            "assignment_id":  a.id,
+            "location_id":    a.location_id,
+            "location_name":  loc.name if loc else None,
+            "project_id":     a.project_id,
+            "project_name":   proj.name if proj else None,
+            "assigned_at":    _iso_utc(a.assigned_at),
+            "assigned_until": _iso_utc(a.assigned_until),
+        }
+
+    # 2. Fetch all events for this serial in one shot.
+    async with httpx.AsyncClient(timeout=_SFM_TIMEOUT_SECONDS) as client:
+        events = await _fetch_events_for_serial(
+            client,
+            serial=unit_id,
+            from_dt=from_dt or datetime(1970, 1, 1),
+            to_dt=to_dt or now,
+            false_trigger=false_trigger,
+            limit=_SFM_FETCH_CEILING,
+        )
+
+    # 3. For each event, walk the assignment list and find the first
+    #    overlapping window.  O(N * M) but both are small in practice.
+    for ev in events:
+        ts_str = ev.get("timestamp")
+        if not ts_str:
+            ev["attribution"] = None
+            continue
+        try:
+            # SFM returns ISO with "T" separator; tolerate both.
+            ts = datetime.fromisoformat(ts_str.replace(" ", "T"))
+        except ValueError:
+            ev["attribution"] = None
+            continue
+
+        matched: Optional[UnitAssignment] = None
+        for a in assignments:
+            a_end = a.assigned_until or now
+            if a.assigned_at <= ts <= a_end:
+                matched = a
+                break
+
+        if matched is not None:
+            ev["attribution"] = _attr_dict(matched)
+        else:
+            ev["attribution"] = None
+            # Find the nearest assignment (chronologically) for diagnostic.
+            if assignments:
+                nearest = min(
+                    assignments,
+                    key=lambda a: min(
+                        abs((ts - a.assigned_at).total_seconds()),
+                        abs((ts - (a.assigned_until or now)).total_seconds()),
+                    ),
+                )
+                # Signed delta in days from the nearest boundary
+                # (negative = event BEFORE that boundary).
+                if ts < nearest.assigned_at:
+                    delta_seconds = (ts - nearest.assigned_at).total_seconds()
+                elif ts > (nearest.assigned_until or now):
+                    delta_seconds = (ts - (nearest.assigned_until or now)).total_seconds()
+                else:
+                    delta_seconds = 0
+                ev["nearest_assignment"] = {
+                    **_attr_dict(nearest),
+                    "delta_days": round(delta_seconds / 86400, 1),
+                }
+
+    # 4. Apply bucket filter.
+    if bucket == "attributed":
+        filtered = [e for e in events if e.get("attribution") is not None]
+    elif bucket == "unattributed":
+        filtered = [e for e in events if e.get("attribution") is None]
+    else:
+        filtered = events
+
+    filtered.sort(key=lambda e: e.get("timestamp") or "", reverse=True)
+    total_count = len(filtered)
+    capped = filtered[:limit]
+
+    # 5. Stats: compute over the ENTIRE event set (not the filtered bucket)
+    #    so the unattributed_count tile is always meaningful regardless of
+    #    which bucket the operator has selected.
+    base_stats = _compute_stats(events)
+    unattributed_count = sum(
+        1 for e in events if e.get("attribution") is None
+    )
+    base_stats["unattributed_count"] = unattributed_count
+
+    return {
+        "events":             capped,
+        "count":              total_count,
+        "stats":              base_stats,
+        "assignments_total":  len(assignments),
+    }
+
+
 # ── Stats helpers ─────────────────────────────────────────────────────────────