Decide bounded-staleness contract for parallel read connections

## Question

Is bounded read staleness acceptable for ActivityWatch read endpoints when reads are served from parallel read-only SQLite connections? If yes, what bounds should we design for?

Erik's current preference is:

- under 1 second in the average case
- never more than 1 minute in the worst case

## Why this gates the design

WAL landed in #616, so readers no longer need to block the serial writer. The remaining fork for the parallel-read TODO in `aw-datastore/src/worker.rs` is consistency:

- if bounded staleness is acceptable, reads can use a read-only pool with that freshness contract;
- if read-your-writes is required, dispatch needs a synchronization/guard path before serving a read.

This issue is only to settle the consistency contract; implementation and benchmarking can follow separately.