chat(): maxIterations doesn't bound tool calls — one turn can fan out unbounded parallel calls

[tombeckenham]

Problem

maxIterations reads like a step/tool-call budget but counts model turns, not tool calls. A single turn can emit unbounded parallel tool calls, so the budget never trips.

Repro

agentLoopStrategy: combineStrategies([
  maxIterations(20),
  untilFinishReason(['stop']),
])

Expected ~20 tool calls. Got ~160, blowing a 300s serverless timeout. The model emitted ~160 calls across a few turns; iterationCount only saw a handful. ~99% of the wall clock was the model generating the fat multi-call turns — tool execution was negligible.

Root cause (3 layers)

1. iterations ≠ tool calls. One assistant turn can carry an array of parallel tool calls; the loop counts turns.
2. No tool-call count in strategy state. AgentLoopState exposes only { iterationCount, messages, finishReason } (packages/ai/src/types.ts:833) — so maxToolCalls is unbuildable in userland.
3. Budget checked only between turns. shouldContinue() runs the strategy at turn boundaries (packages/ai/src/activities/chat/index.ts:1936); a single runaway turn is never inspected mid-flight.

Proposed

• Add cumulative toolCallCount (and per-turn lastTurnToolCallCount) to AgentLoopState.
• Ship a maxToolCalls(n) strategy.
• Cap parallel fan-out per turn (bound calls executed in one turn — the direct fix for the 160-in-one-turn case).
• Optional: a deadline(ms) / wall-clock-aware strategy, since call-counting alone can't catch a single slow turn.
• Docs note: iterations ≠ tool calls when parallel tool use is enabled.

Minimum to close the report: AgentLoopState.toolCallCount + per-turn fan-out cap. Needs E2E coverage (fixture emitting a fat parallel-tool-call turn) + docs update per repo convention.

Seen on @tanstack/ai@0.28.0.