Fix duplicate verbal responses: send response.create ONCE per turn, not per tool

Add rim pricing rule: single rim price, specify front/rear when quoting

Use exact English pronunciation rules text as user provided

Add tool search rules + Caddebostan pronunciation to prompts.py (kalici)

Extend full-name rule to jant/gidon/pedal/lastik etc with category-aware examples

Force GPT to pass full product names to tool (avoid fuzzy mismatches)

Strengthen: ANY product mention triggers tool call (sele, kask, far, all categories)

Allow short filler phrase before tool call (stock search takes 2-3s)

Hard cap: max 2 sentences (no 3rd sentence)

Strict ban on closing/filler phrases (mağazada inceleyebilirsiniz vb.)

Prompt: call get_warehouse_stock whenever mentioning a specific product (for image display)

Revert to gpt-5.5 (nano missed product matches)

Switch fuzzy matcher to gpt-5-nano (10x cheaper, ~2s faster)

Trust GPT for product selection: only use tool's productLink, drop transcript-based guessing

Drop '?' from question_indicators — was filtering 'Marlin 5 var mı?' as non-product

Trek PHP primary (proven cache+stale), BizimHesap direkt fallback

Drop Trek PHP fallback — BizimHesap direkt only (single source of truth)

Trek PHP timeout 8s -> 3s (endpoint often down — fail fast, use BizimHesap)

Route smart_warehouse cache through V2 lock'lu stock_service+product_index (no more thundering herd)

Use WhatsApp BF-WAB smart_warehouse_with_price + prompts (proven quality)

Speed: local matcher primary; tukendi message when catalog hits but no stock

Restore gpt-5.5 as primary fuzzy matcher (V1 quality), local + nano as fallback

Use websockets modern asyncio client (legacy fails with uvloop)

Initial: BF-Realtime V2 clean architecture