Remove corpus counts from chat chrome

apps/ai-service/rag/budget.py

@@ -1,21 +1,20 @@
 """A per-request LLM-call budget — F-08.
 
 `RagAgent.handle()` makes up to ~5 sequential Bedrock calls per turn
-(understand, sufficiency, generate, up to 2 entailment retries) with no
-aggregate deadline before this: each call is bounded only by its own fixed
-provider timeout (`read_timeout=60` in `adapters/bedrock_converse.py`, times
-up to 3 retries at "standard" backoff — worst case several minutes for one
-stuck call, let alone five). Measured live 2026-08-07: a normal answerable
+(understand, sufficiency, generate, entailment) with no aggregate deadline
+before this. Each call now has at most two bounded provider attempts
+(`read_timeout=20` in `adapters/bedrock_converse.py`) because this budget is
+checked between calls and cannot cancel boto3 while it is already in flight.
+Measured live 2026-08-07: a normal answerable
 turn costs ~8-9s total; nothing bounds the pathological case.
 
 Checked before each call, not wrapped around an already-running one — this
 bounds how many MORE calls get a chance to start once time/calls run out. It
 does not cancel a call already in flight past its own provider timeout; a
 hard per-call cancellation would need cooperative cancellation support from
-`adapters/bedrock_converse.py`'s boto3 client, a larger change than this
-budget object alone. Still a real improvement: five calls each capable of
-running to their own 60s+ limit, one after another, is the actual gap this
-closes.
+`adapters/bedrock_converse.py`'s boto3 client. The adapter-level timeout bounds
+that residual gap; this object prevents any later call from starting after
+the aggregate deadline.
 """
 from __future__ import annotations