[Maharsh17]

Each event now feeds a RunDistanceStream, a SessionSegmenter and the goal stream together, in one place, so a recompute reads their state instead of redoing the whole buffer. The output is the same as the batch calls give over the buffer until it wraps. A test checks exactly that.

What changes

• Run indices count from the worker's first event, so they stay stable when the buffer wraps and stay joined to the goal stream. run_count counts every run since the worker started.
• The episodes payload drops the per event list that nothing reads. Episodes and pauses keep the same window the buffer holds.
• Stored goal runs are released from the stream, except the latest, which a playgroundData outcome may still re-profile.
• Stable run numbers across backfills and restarts; Reset clears the whole student tile #97's session keying is kept as is, and Fix bugs from the repo audit: event order, stale data after untrack, reset/export gaps #98's oldest first ordering is what the streams rely on.

This bumps the engine to f03612b, which is main again with the scaling fixes from agent-lm-packages #8. The humanize and hat tests are updated for the engine's explicit hat list and its not connected section.

All 267 tests pass, including test_run_identity and test_audit_fixes.

Closes #99