fix(ticketing): wire alarm and grid-state facts into alert judgment

[1vav]

Summary

The live alert-correlation LLM call writes every alert's "Likely cause" and
"Suggested action" lines and decides whether a new alert joins an existing
ticket. It never saw VRM alarm data or grid-state duration facts, even
though the prompt it's given explicitly instructs it to use both. Both now
reach the prompt.

These ship together because they're the same mistake in two places: when
the correlation prompt was rewritten around a new typed context object
(AlertJudgmentContext), two sources the old prompt fed the model had no
matching field in the new schema and were silently dropped rather than
erroring.

What was missing

• VRM alarm / power-trend evidence. The fetch is real (VRM's actual
  alarm-log and diagnostics endpoints, scoped to the right site and a
  30-minute window) and the raw data reaches the context-assembly layer —
  but AlertTelemetry, the schema that shapes what the prompt actually
  sees, never declared the three fields, so pydantic's extra="ignore"
  dropped them before the LLM call. The prompt names "overload" as the
  literal example of what this evidence should surface.
• Grid-state duration facts (is_hps_on / is_hps_on_updated_at / DCU
  status roll-up). The prompt's "Root Cause Rules" section — the mechanism
  for deciding whether a new alert is a child of an existing grid-off state
  — depends on this, and the pre-rewrite prompt path fetched it. The
  rewritten path never called the fetch at all.

Fix

• Add the three alarm/power-trend fields to AlertTelemetry, matching the
  names the prompt already expects, with a small alias step mapping the
  raw provider's own key names onto them before validation — so neither
  the provider nor the prompt vocabulary has to change, only the one seam
  between them.
• Add a grid_operational_facts source to AlertJudgmentContext, gathered
  by the assembler the same resilient way every other source already is
  (bounded timeout, failure isolation, availability tracking). Left as a
  raw dict rather than a new narrow sub-model, since a stricter schema is
  exactly what dropped the alarm fields the first time.
• Wire the new provider at the one real call site and add the
  corresponding prompt section.

Testing

Extended the assembler and prompt-builder unit tests to pin both sources
surviving end to end — raw provider data in, correct field names in the
built prompt out — plus the EMPTY-vs-FAILED distinction for the new
grid-facts source (a lookup miss is "no extra context," not degradation).
Full chat_orchestrator suite and pre-commit run --all-files clean.

---