feat(rag): Phase 1 — rewrite build_card_text to natural-language format (#137)

* feat(rag): Phase 1 — rewrite build_card_text to natural-language format

* feat(rag): Phase 2 — Forge-enriched ChromaDB RAG store (2a–2e)

Phase 2a: Forge lazy import guard (_try_import_forge_lookup)
Phase 2b: _build_document appends Forge trigger/replacement/static/keyword tags
Phase 2c: _build_metadata adds has_trigger, has_replacement, has_static, trigger_modes fields
Phase 2d: query_cards extended with require_trigger, require_replacement, require_static params + ChromaDB where clause support
Phase 2e: get_rag_stats reports forge_enriched and forge_coverage; _write_build_meta tracks coverage; build loop counts enriched cards

* feat(rag): Phase 3 — add mtgrules second ChromaDB collection

Adds build_rules_index() and query_rules() operating on a separate
'mtgrules' ChromaDB collection. Chunks MagicCompRules.txt by rule-number
boundaries, embeds each chunk, and exposes semantic search. Completely
independent from the mtg_oracle card collection.

* feat(coach): Phase 4 — behavior-aware RAG + rules context in coach_service

Phase 4a: run_coaching_session uses dual RAG query
  - Primary: commander synergy (n=10)
  - Secondary: require_replacement=True for graveyard/protection cards (n=5)
  - Deduplicates by name before merging
Phase 4b: query_rules() fetches 3 Commander rules chunks to ground the prompt
  - Passed as rules_context kwarg to build_user_prompt()

* feat(rag): Phase 5 — add /api/rag/build-rules and /api/rag/query-rules endpoints

services/card_text.py

@@ -3,35 +3,43 @@ services/card_text.py
 =====================
 Shared utility for building card text representations.
 
-Both services/rag_store.py and coach/embeddings.py need to produce
-identical text strings for embedding so that vector similarity is
-consistent across both systems.  This module provides a single
-canonical implementation.
+    Both services/rag_store.py and coach/embeddings.py need to produce
+    identical text strings for embedding so that vector similarity is
+    consistent across both systems.  This module provides a single
+    canonical implementation.
 
-Format:  Name | Mana Cost | Type | Oracle Text
-(fields that are empty/missing are omitted)
+    Format:  Natural-language sentences.
+    (fields that are empty/missing are omitted)
 """
+
 from __future__ import annotations
 
 
 def build_card_text(card: dict) -> str:
-    """Build the canonical text representation of a card for embedding.
+    """Build canonical natural-language card text for embedding.
+
+    Produces '{Name} is a {Type}. It costs {ManaCost}. {OracleText}'
+    Fields that are empty/None are omitted.
 
     Args:
         card: Dict with keys 'name', 'mana_cost', 'type_line', 'oracle_text'.
               Missing keys are treated as empty strings.
 
     Returns:
-        Pipe-separated string like 'Sol Ring | {1} | Artifact | {T}: Add {C}{C}.'
+        Natural language string suitable for nomic-embed-text or GTE models.
     """
-    parts = [card.get("name", "") or ""]
-    mana_cost = card.get("mana_cost", "") or ""
-    if mana_cost:
-        parts.append(mana_cost)
-    type_line = card.get("type_line", "") or ""
-    if type_line:
-        parts.append(type_line)
-    oracle = card.get("oracle_text", "") or ""
+    name = (card.get("name", "") or "").strip()
+    mana = (card.get("mana_cost", "") or "").strip()
+    types = (card.get("type_line", "") or "").strip()
+    oracle = (card.get("oracle_text", "") or "").strip()
+
+    parts: list[str] = []
+    if name:
+        base = f"{name} is a {types}." if types else f"{name}."
+        parts.append(base)
+    if mana:
+        parts.append(f"It costs {mana}.")
     if oracle:
         parts.append(oracle)
-    return " | ".join(parts)
+
+    return " ".join(parts) if parts else name