fix(agent): answer sibling tool calls when the exit validator rejects finish

[wolfy-j]

Symptom

When an assistant turn contains the finish exit call plus other tool calls (e.g. parallel KB searches), and arena_config.exit_func_id rejects the finish (missing declared outputs), the workflow permanently stalls on the next iteration with:

messages.N: `tool_use` ids were found without `tool_result` blocks immediately after: toolu-...

Observed in production: a digest whose exit validator enforces complete outputs failed every scheduled run for four days with exactly this error.

Root cause

node/agent/node.lua, process_tool_results: on validator rejection the code records the rejection observation, sets skip_call = true, and returns early via if task_complete or skip_call then return ... end — skipping the loop below that records tool_result observations for the sibling tool calls in the same turn. Those siblings already executed (their output sits in tool_results); they only lack recorded observations. The next prompt rebuild therefore carries orphaned tool_use blocks and the LLM provider rejects the whole request.

Fix

Drop skip_call; gate the early return on task_complete alone, so a rejected finish falls through into the existing sibling-observation loop. The loop already no-ops for the exit call itself (it is never in tool_results — split_exit_tool_calls keeps it out of execute_tools), so nothing is double-recorded, and the genuine-completion path is unchanged.

Tests

process_tool_results exposed via the existing _test table; new process_tool_results_test.lua with a stub always-rejecting exit validator:

1. rejected finish + two sibling calls in one turn → exactly 3 observations recorded (rejection + both siblings), task_complete == false;
2. unconditional finish (no exit_func_id) → early return, sibling results untouched.

Full suite: 952 passed.