fix(orchestration): dirty re-analysis duplicates max_parallel logic from analyze()

## Context
In `DagScheduler::tick()`, after deferred topology re-analysis (when `topology_dirty=true`), the `max_parallel` calculation is duplicated inline rather than using `TopologyAnalysis.max_parallel` from `analyze()`. These will silently diverge if the formula changes.

## Fix
Use the `max_parallel` value from `TopologyAnalysis` returned by `analyze()` — remove the inline duplication.