Tape parity: limited-read comparison uses the snapshot window

[ReganBell]

The full-corpus parity check compares a session's entries and tape from one snapshot, but the limited-read comparison calls the live transcript source, which re-reads the store — so a turn landing between the script's snapshot and the serving read makes the served view newer than the expected list, and the trailing rows count as real mismatches. On an active fleet this produces a steady band of false reals per run (reproducibly, since the settle-recheck races the same ongoing activity).

Fix: compare only the intersection window — served rows clamped to the snapshot's latest seq, expected entries clamped to the tape's covered seq. A race-faithful pinning test (snapshot, then a live turn, then the limited comparison with the stale snapshot) is red on the old code.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}