feat(bot): dialog engine — scrub → safety → FSM → monitoring, consent, age gate, crisis

- Engine.handle registered via register_engine; lexicon-first safety (no LLM on ACUTE),
  fixed crisis script once per open crisis, POSSIBLE → C-SSRS screen
- FSM registry (@state/@on_enter/@trigger) with auto-loaded flows; core flow:
  new → consent → age_gate → module onboarding → idle, crisis, consent_revoked
- reusable scale session (buttons, extract_item + clarify, item 9 / C-SSRS buttons-only,
  crisis flags), execution of monitoring dialog actions, texts in config/texts/dialog.yaml

Refs #10

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app/dialog/crisis.py

@@ -0,0 +1,49 @@
+"""Crisis path of the dialog (guardrail #2): fixed texts only, never generated by the LLM.
+
+Decisions:
+  * The crisis script (`app.core.safety.crisis_reply`, `config/texts/crisis.yaml`) is sent verbatim, once per
+    open crisis. "Open" = `data["crisis"]["open"]` — set when the crisis starts, cleared only by the doctor's
+    acknowledgement (`doctor_ack`), so a patient repeating the phrase does not get the same long script again:
+    they get a short fixed "я рядом, врач уже уведомлён" + emergency resources instead.
+  * Entering the crisis drops unfinished scale sessions (their answered items are already stored item by item):
+    nothing else is asked while the patient is in crisis, except the C-SSRS screen.
+  * Without valid consent (before consent, after refusal/revocation, minors) we still answer a crisis message —
+    protecting life is the one processing ground that does not need consent (152-FZ art. 10 p. 2.3) — but with
+    a variant WITHOUT "your doctor has been notified", because no doctor is notified, and nothing is stored.
+"""
+
+from typing import Any
+
+from app.core.safety import CrisisResource, CrisisScript, crisis_reply
+from app.dialog.contracts import Reply
+from app.dialog.fsm import Ctx, Step
+
+
+def crisis_open(data: dict[str, Any]) -> bool:
+    return bool((data.get("crisis") or {}).get("open"))
+
+
+def _resource_lines(resources: list[CrisisResource]) -> list[str]:
+    return [f"• {r.name}: {r.phone}" + (f" ({r.note})" if r.note else "") for r in resources]
+
+
+def open_crisis(ctx: Ctx, categories: list[str], source: str) -> Step:
+    """Start the crisis (full script) or, if one is already open, send the short repeat text.
+
+    `source` is a code for the dashboard/audit trail: "safety:lexicon", "safety:llm", "scale:phq9", "rule:<id>".
+    """
+    script = crisis_reply(categories)
+    if crisis_open(ctx.data):
+        text = "\n".join([ctx.texts.crisis.repeat, *_resource_lines(script.resources)])
+        return Step(replies=[Reply(text=text, kind="crisis")])
+    crisis = {"open": True, "opened_at": ctx.now.isoformat(), "source": source, "variant": script.variant}
+    return Step(
+        replies=[Reply(text=script.text, kind="crisis")],
+        next_state="crisis",
+        data={"crisis": crisis, "scales": None},
+    )
+
+
+def no_consent_crisis_text(script: CrisisScript) -> str:
+    """Crisis script without the "doctor notified" line — for people we may not (or did not) notify about."""
+    return "\n".join([script.body, "", *_resource_lines(script.resources), "", script.grounding])