fix(reminders): remove execution cap so reminders never skip on capacity

[Aaronontheweb]

Problem

The daily vendor-price audit was skipped this morning (Aug 9). Root cause chain:

1. Capacity gate (MaxConcurrentExecutions = 3, hardcoded) was reached — three tbench/PR test reminders were parked on long executions burning their full 1h timeouts.
2. PR fix(reminders): retain failed one-shot executions #1812 changed the capacity path from defer-and-redispatch to settle-blocked-occurrence — it acked the occurrence as skipped instead of queueing it.
3. That settlement ack raced with the cron reschedule (manager re-arms tomorrow's occurrence before acking today's), so the ack returned NotFound, which tripped a reminder.settlement.failed ops alert for a scenario that is not an ops failure.

Fix

Remove the execution cap entirely:

• Deleted MaxConcurrentExecutions const and the capacity gate in ReminderManagerActor.
• Reminders now always start executing immediately — no capacity skip, no settlement, no alert noise.
• Failure/retry semantics are unchanged: Akka.Reminders owns Nack/retry, and every execution still has a one-hour absolute timeout (ExecutionAttemptTimeout), so unbounded scheduling pressure on the LLM is acceptable.
• Updated the scheduling spec (base + fix(reminders): retain failed one-shot executions #1812 change spec) to match: manager SHALL NOT cap concurrent executions.

Tests

• Reworked Recurring_occurrence_at_capacity_is_acked_without_execution into Recurring_occurrence_starts_even_while_other_reminders_are_running: three blockers in flight, a fourth reminder fires — asserts it is dispatched (0 skipped duplicates) and executed by the pipeline.
• Full ReminderManagerActorTests suite passes: 35/35.

Notes

• The interval in the new test was bumped to 2h so the occurrence passes the pre-existing safe-execution-lease check (deadline = due + interval must exceed 1h+1m); the old test never reached that path because the capacity gate short-circuited first.

[Aaronontheweb]

Now that we tolerate longer timeouts for reminders, we can afford to apply more scheduling pressure even to smaller, constrained local LLMs. the hard limit of 3 concurrent reminders + buffering we did to the rest is effectively obsolete now