fix: TECH-1717 agent & chat reliability fixes + smoke test coverage - #613

• Retry — ExecutionTimeoutError is now retried; GuardrailValidationError warns before re-raise; an explicit per-call retry= overrides the agent-level setting (previously unreachable dead code).
• Concurrency — per-run state (run_id / output / tool counts) isolated in a ContextVar (agent/run_state.py); concurrent do_async/astream on one agent no longer clobber each other.
• Chat session state — a failed invoke latches SessionState.ERROR; reopen() recovers non-destructively (history + cumulative duration); invoking a closed session raises SessionClosedError.
• Usage — tool_calls are counted into the usage registry (from ModelResponse.tool_calls).
• Memory — num_last_messages groups history by user-turn boundary, so the limit holds through tool round-trips.
• Heartbeat — execute_heartbeat / aexecute_heartbeat gain a print= opt-in (silent by default).
• Refactor — reuse the ModelResponse.tool_calls property instead of inline ToolCallPart filtering (recorder + 3 call sites).

Usage-metrics accuracy

Three pre-existing telemetry/display defects (low severity, no wrong outputs), fixed and regression-tested:

• Pipeline step count — MemorySaveStep was the only non-streaming step that never self-finalized, so runs logged 23/24 steps and dropped step_statuses['memory_save']. It now finalizes in a finally like every sibling → 24/24.
• Tool execution time — tool wall time was folded only into the per-run snapshot, so registry-derived Agent.usage.tool_execution_time stayed 0. Now recorded as a kind="tool" registry entry via record_tool_execution_time, through the single AgentRunOutput.add_tool_execution_time chokepoint (no double-count). upsonic_execution_time is unchanged (the same elapsed is added to both duration and tool_execution_time).
• Cached cost — the per-call cost panels priced cached input at full rate, over-estimating vs the cache-aware chat.usage.cost. Panels now price cache-aware via get_estimated_cost_from_usage (backward-compatible: no cache tokens → identical output).