[Blind-Striker]

Tests only, one file. Backs a claim #59 made canonical: a v2.session.log cursor past the aggregate's tail is accepted rather than refused, and then suppresses live delivery until the aggregate overtakes it. That came from reading the pinned server, and nothing measured it. Upstream has no production caller that passes after at all, so this SDK is the path's first consumer and the behaviour was unexercised on both sides.

The unreachable cursor. A cursor a thousand past the tail is read to the end. The reply is the marker alone, and the marker carries the real watermark rather than the cursor it was asked for. Its control is the settled read taken first, which returns the turn's own terminal event, so an empty answer is the cursor being out of range and not a server that retained nothing. Persistence is a server option, so without that control this proof would pass unchanged against a server that had it switched off.

Skip and resume. The second proof puts the cursor three past the tail, where the following turn crosses it, so one attachment shows both halves. The execution start the cursor covers never arrives; the terminal event beyond it does. A second reader on the correct cursor returns that skipped start from the same session, so its absence on the attachment is a skip rather than a gap, and the resumed terminal event is what proves the attachment was alive rather than quietly dead. Silence is never the evidence here: every assertion is something that did arrive, somewhere.

The recorded run: watermark 10, attachment cursor 13, skipped execution start at sequence 12, delivery resuming through the terminal event at 19.

Evidence. Full Release suite: 5022 tests, 0 failed, 0 skipped, nine assemblies across net8.0, net9.0, net10.0 and net472. generate --verify and refresh-spec --verify pass on the same tree.