After a mid-turn cost correction the model launched a thirty-call MCP batch instead of pacing or confirming spend

**What happened**
After the user's mid-turn message flagged budget waste and named the graph tools to use, the model issued about thirty MCP calls in one batch (6 tool_map, 6 cluster reads, 11 cypher, 6 context, 6 trace). tool_map returned no tools for four of six repositories and five of six traces returned no_path, ambiguous or truncated.

**Expected**
A small first batch to learn which graph tools return data for these indexes, and a pause to confirm further spend.

**What the user said**
"you stole from me."

**Repro**
Correct the model on tool choice and cost mid-turn; observe the size of its next batch.

**Environment:** Claude Code CLI 2.1.270 (Claude Code), macOS, auto mode, model claude-fable-5-1, session of 2026-09-14.

Start by reproducing the stated mid-turn correction in Claude Code CLI 2.1.270 with MCP enabled, then observe the next batch size and tool results. Compare the behavior with the expected small exploratory batch and spend-confirmation pause; done means the correction no longer leads to an approximately thirty-call batch without pacing or confirmation.

Written by the indexing model from the issue text.