Usage events are counted once, keyed by the full session line

Forked and cloned sessions copy their ancestor's entry lines byte-identical into the new file. A scan that sums every file would count the shared history once per fork. Each usage event is therefore counted once per distinct line: the scan keeps the set of full JSONL lines that carried a usage event and counts a line only the first time it appears.

Considered options

• Entry id. Entry ids are 8 hex chars and random; they collide across unrelated sessions, so an id alone cannot tell a copy from a collision.
• The parentSession header. It names one parent, not the whole copied chain, and older forks may lack it.
• (session id, entry id) pairs. A fork is a different session, so the pair does not merge the copy with the original.

Consequences

• Counting depends on pi writing fork copies byte-identical. If that ever changes, the dedupe silently overcounts instead of undercounting; a test on a fork fixture guards the invariant.
• Two truly distinct events with identical bytes would merge. In practice an event line carries a random entry id, so identical bytes mean a copy.