Add config knob to throttle background summarizer speed #71

Summary

The summarizer's enqueue_catch_up() dumps all unsummarized turns into the queue at once and processes them as fast as inference allows. When a large batch of historical turns is imported (e.g. from a Claude.ai export), this can monopolize the Beelink inference endpoint for hours.

Proposed config (in [housekeeping] or [summarizer] section of settings.ini)

summarizer_delay_ms = 0   # ms to sleep between summarizer jobs (0 = no throttle, default)

Implementation

In SummarizerService's worker loop (src/summarizer_service.cpp), after completing a job and before picking up the next, sleep for config().summarizer_delay_ms when the value is > 0. This keeps the queue draining but at a user-controlled pace.

Use cases

• Normal operation: 0 (no delay, drain as fast as possible)
• Running alongside active inference workloads: e.g. 5000 (5s between jobs ≈ ~12/min)
• Overnight catch-up import: leave at 0 to finish fast

Notes

• The delay applies per-job (L2, L3, L4, draft retry) — simplest and most predictable
• SIGHUP-reloadable so it can be adjusted without a daemon restart
• Pairs naturally with the existing housekeeping_interval cadence