feat(matching): Jev entity matching phases 2-5 (per-field fan-out, variants, candidate ambiguity buttons)

Phase 2 — jev.py client: POST /v1/systemone, PINNED model jev-1.13.0 (drift warning when the response model differs), one retry on 429/5xx, fail-soft (None on any error), usage/latency returned for observability. Config adds TYPESAFE_API_KEY / JEV_* (thresholds 0.85/0.60, 15 questions per field); no daily cap by owner decision. bot/.env is now mounted read-only into the container so the key never enters compose/git.

Phase 3 — matching.py: برون‌گروهی (one request per field, parallel), درون‌گروهی (one choice question per mention sharing the same full candidate list + «_none»), درون‌واریانسی (all STT spellings in one question). Routing: conf>=exact → bind; floor..exact → ambiguity; else/_none → new; missing/failed field → difflib fallback (bug fixed during testing: a failed request used to record 'new' instead of falling back). Resolver gains full cached lists for suppliers/materials (paged) plus brands (API+DB) and specs (EAV values + draft specs). NLU prompt now emits entity_mentions (6 fields) with the voice-variant grouping rule; sanitize_mentions guards the shape. handle_message runs matching only for register_quote/update_draft/add_material behind JEV_ENABLED; jev_map threads into quote/material resolve_payload (Jev hit skips difflib; exact writes ids; ambiguity skips difflib). Spec effect per owner: spelling normalization only (normalize_brand_spec).

Phase 4 — entity-match ambiguity sub-messages: create_ambiguities() builds deduped entries {kind: entity_match, field, variants, candidates, details with Persian percentages}; new keyboards.ambiguity_candidates_keyboard renders one button per candidate + «هیچ‌کدام — ایجاد جدید»; callbacks handle amb⛏️{prefix}:{amb_id}:{idx} (applies the candidate via matching.apply_choice then re-resolves/renders) and amb:new (rejected → auto-create path at commit). Every user choice lands in observability (ambiguity.user_choice with arg+candidates).

Phase 5 — voice variants: NLU groups different STT spellings of one entity into one variants[] (verified live: مصطفوی/مستفدی → one supplier mention); tools/test_matching.py live section (RUN_LIVE=1) exercises real DeepSeek + real Jev without touching Bale.

Tests: test_jev.py 5/5 (mock HTTP: no-key, happy path, 429-retry, 400, exception), test_matching.py 11/11 mock + 3/3 live (real NLU produced 7 mentions with correct fields; real Jev matched «50 کیلویی»→«۵۰ کیلویی» 0.98 and «کیسه» 1.0 and correctly returned new against the emptied catalog); observability regression 8/8.

bot/app/nlu.py

@@ -25,6 +25,23 @@ def normalize_digits(text: str) -> str:
 
 ZWNJ = "\u200c"
 
+MENTION_FIELDS = {"project", "supplier", "material", "brand", "spec", "unit"}
+
+
+def sanitize_mentions(raw) -> list[dict]:
+    """Keep only valid {field, variants} mention objects from the LLM output."""
+    out: list[dict] = []
+    if not isinstance(raw, list):
+        return out
+    for m in raw:
+        if not isinstance(m, dict):
+            continue
+        field = str(m.get("field") or "").strip()
+        variants = [str(v).strip() for v in (m.get("variants") or []) if str(v or "").strip()]
+        if field in MENTION_FIELDS and variants:
+            out.append({"field": field, "variants": variants[:6]})
+    return out
+
 
 def normalize_text(text: str) -> str:
     """Normalize Persian text for fuzzy matching (ZWNJ, arabic ya/ka, spaces)."""
@@ -60,6 +77,7 @@ SYSTEM_PROMPT = """تو دستیار هوشمند تدارکات یک سیستم
   "consensus_text": "بیانیه رسمی، پالایش‌شده، فشرده و بدون تکیه‌کلام از تقاضای تدارکات کاربر",
   "intent": یکی از {intents},
   "patch": {شکل partial هر فعالیتی که کاربر داده},
+  "entity_mentions": [{"field": "project|supplier|material|brand|spec|unit", "variants": ["همه‌ی نگارش‌ها"]}],
   "confidence": 0.0..1.0
 }
 
@@ -95,6 +113,15 @@ SYSTEM_PROMPT = """تو دستیار هوشمند تدارکات یک سیستم
 - اگر درباره پذیرش استعلام است → intent="accept_quote" و در patch.target نام تأمین‌کننده/استعلام.
 - اگر درباره اعلام رسیدن اقلام است → intent="arrive_items" و در patch.target نام تأمین‌کننده/استعلام و در patch.arrivals مقادیر.
 
+۶. استخراج موجودیت‌های قابل تطبیق (entity_mentions):
+   - فیلدهای مجاز: "project" | "supplier" | "material" | "brand" | "spec" | "unit"
+   - هر نام کالا/تأمین‌کننده/پروژه/برند/واحد و هر مشخصه که در پیام آمده، یک عضو جداگانه است.
+   - اگر در نسخه‌های مختلف STT همان نام با املاهای متفاوت پیاده شده (مثل «مصطفوی»، «مستفدی»، «مسطفی»)،
+     همه را در variants همان یک عضو بگذار — نه چند عضو جداگانه.
+   - برای پیام متنی معمولی، variants یک عضو تک‌عنصر است.
+   - مشخصه (spec) و برند (brand) جدا از متریال و جدا از هم استخراج شوند.
+   - وظیفه‌ی تطبیق با موجودیت‌های سیستم بر عهده‌ی تو نیست؛ هیچ نامی را به نزدیک‌ترین نام موجود «حدس» نزن.
+
 شکل patch برای intent="register_quote"/"update_draft":
 {"supplier_name": str, "project_name": str, "project_names": [str], "is_official": bool, "notes": str, "expiry_days": int, "expiry_date": "YYYY-MM-DD",
  "items": [{"material_name": str, "brand_name": str, "specs": [str], "quantity": number, "unit_name": str, "unit_price": number, "project_name": str, "notes": str}],
@@ -148,7 +175,8 @@ def _extract_json(raw: str) -> dict | None:
 async def understand(settings: Settings, text: str | list[dict], draft_summary: str | None = None,
                      trace_id: str | None = None) -> dict:
     """Return {"consensus_text": ..., "intent": ..., "patch": {...}, "confidence": ...}; never raises."""
-    fallback = {"consensus_text": "", "intent": "unknown", "patch": {}, "confidence": 0.0}
+    fallback = {"consensus_text": "", "intent": "unknown", "patch": {},
+                "entity_mentions": [], "confidence": 0.0}
     if not settings.AI_BASE_URL or not settings.AI_API_KEY:
         logger.warning("AI provider not configured; cannot understand message")
         return fallback
@@ -214,6 +242,7 @@ async def understand(settings: Settings, text: str | list[dict], draft_summary:
                             payload={"error": "unparseable", "raw": content[:500]})
             return fallback
         parsed.setdefault("patch", {})
+        parsed["entity_mentions"] = sanitize_mentions(parsed.get("entity_mentions"))
         if not parsed.get("consensus_text"):
             parsed["consensus_text"] = first_text
         await obs.event(trace_id, "nlu.result", ok=True, model=settings.LLM_MODEL, payload={
@@ -221,6 +250,7 @@ async def understand(settings: Settings, text: str | list[dict], draft_summary:
             "confidence": parsed.get("confidence"),
             "patch": parsed.get("patch"),
             "consensus_text": parsed.get("consensus_text"),
+            "entity_mentions": parsed.get("entity_mentions"),
         })
         return parsed
     except (httpx.HTTPError, json.JSONDecodeError, KeyError, TypeError):