ATperson archive replay fixes and wolfram 0.23.x

The year-back Jetstream archive backfill (#144) was blocked by three archive replay bugs and two upstream wolfram bugs. All fixed; the backfill now sweeps 10M-sequence windows from zero to the sealed tip (26,292,300,745).

Changes

• Archive host binding — JetstreamReplayClient reused the agent's PDS XRPC client, so every plan call hit bsky.social instead of the archive API. The client now owns its own wf_xrpc_client bound to ATPERSON_JETSTREAM_ARCHIVE_HOST (default https://jetstream.us-west.bsky.network). No PDS session needed for replay.
• Sealed-tip probe — the archive clamps sealedTipSeq to the request's beforeSeq, so the old probe (before=1) always returned 1. The probe now sends afterSeq beyond the tip with no upper bound; the response carries the unclamped global tip.
• Exhaustion signal — fetch_window reported the clamped window end as the sealed tip, so every auto-capped window claimed "sealed archive exhausted" after one run. The window probes the true tip once and uses it for both the clamp and the exhaustion signal.
• wolfram 0.23.0 — replay plan seq values serialise as exact JSON integers. cJSON's double path emits scientific notation past ~1e15, which the archive rejects with a 400.
• wolfram 0.23.1 — the XRPC transport follows HTTP redirects (bounded, 5 max). The archive's getBlock 307s to a CDN URL; without CURLOPT_FOLLOWLOCATION every replayed block fetch failed.