Config: YAML booleans silently collapse the subagent batch scheduler's numeric knobs (attempts, running items, caps) #6290

Summary

SubagentBatchesConfig gives the durable subagent batch scheduler its own numeric knobs, but they accept YAML booleans. Pydantic coerces true to 1 before the constraint check, so with:

subagent_batches:
  enabled: true
  max_attempts: true             # intended 3 - loads as 1
  default_max_running_items: true # intended 3 - loads as 1
  max_running_items_per_batch: true # intended 64 - loads as 1

...the config loads cleanly and, since the batch service captures these limits once at Gateway startup:

• max_attempts: true -> every batch item gets a single attempt, so any transient failure is terminal;
• default_max_running_items: true and max_running_items_per_batch: true -> batch execution serializes to one running item;
• max_items_per_batch: true, default_max_live_items: true and max_live_items_per_batch: true -> the batch/live-item caps collapse to 1;
• poll_interval_seconds: true -> loads as 1.0 on the float field - the same value as the default today, but a mistyped boolean silently accepted as a number.

Verified live on main (pydantic 2.13.3): the nine int fields accept True as 1; lease_seconds, max_result_chars and result_preview_max_chars reject it only via their higher ge floors, with a misleading "greater than or equal to" message.

Fix

Mirror the scheduler-config boolean guard (#6258): register the shared boolean guard (#6226) as before-validators rejecting booleans with a field-specific error across all ten numeric fields of SubagentBatchesConfig (split by wording: an integer for the int fields, a number for the float poll interval, the way the extensions config already does). PR follows immediately.