Every factual claim in this file verified out: both endpoints are served, the cursor magnitude constant is 1_000_000_000_000_000 in serve/server.zig, the v1 dict id 1612007021 and the v2 id 20260709 at 65,536 bytes are all real and the v2 pair is test-asserted down to a SHA-256. Nothing corrected; what was missing is a way in. The file uses "v1" and "v2" on nearly every line without ever saying how they relate or which Stream serves (both). Added a two-row table for the only real differences -- compression opt-in and the optional cursor envelope field -- and spelled out that "v1 quirk" means deliberately preserved bug-compatible behaviour, not a defect awaiting a fix, which is the whole point of the checklist at the end. Explains the cursor magnitude split, which reads as arbitrary: microsecond timestamps are already ~1.7e15 and sequence numbers start at 1, so the ranges cannot collide until one instance assigns 10^15 events. Records that v2_dictionary_id is derived from the embedded dictionary's own header, so the advertised ID and served bytes cannot disagree -- and that grepping for 20260709 finds nothing because Zig writes it 20_260_709. That digit-separator false negative cost a detour on this pass. Marks the upstream Go source paths and the docs/README.md 3.5 citation as upstream, consistent with the other two upstream specs. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>