Orchestrator — Mid-Task Skill Injection

Tracks active domain keywords as the task evolves. On a sustained domain shift (cluster of 3+ new keywords that don't match the active domain), proposes loading a relevant skill via a pragya checkpoint. Never auto-loads — the human always decides. At most two injection suggestions per task to avoid checkpoint fatigue. Declined suggestions are not re-raised unless the context deepens further. If disabled, librarian and pragya still work; only mid-task injection is lost.