P1b: the frontdoor flips to source-side sequencing — drop the ring, pass epoch:seq ids through, the lossless-resume E2E acceptance #1813

Problem — P1a (merged to the lane) put the sequence at the source: the engine stamps :, the shared 32MB buffer replays cursors, unservable cursors reset. The frontdoor still runs its own 512-frame ring (~8s of history at flood rates) with a 64-frame unknown-id replay cap — the reload window that expires under load and misses running sessions. This is P1b of #1803: the frontdoor demotes to a dumb pipe.

Approach — The frontdoor's merged event groups pass the engine's id lines through untouched; its ring, its id re-stamping, and its replay logic retire. The engine's replay (via the frontdoor forwarding Last-Event-ID) serves reconnects. The lossless-resume E2E is the acceptance: kill the link mid-turn, reconnect, transcript equals fresh load — run against the sharded hub.

Acceptance Criteria

• The frontdoor forwards the upstream's SSE id lines verbatim to members (epoch:seq on the wire end to end).
• The member's Last-Event-ID reaches the engine upstream (the frontdoor no longer terminates cursor semantics).
• The ring code and its HISTORY_MAX/REPLAY_CAP constants retire (or reduce to a documented passthrough); the frontdoor's replay tests retire with it.
• The lossless-resume E2E passes against the sharded hub: mid-turn kill, reconnect, the transcript matches a fresh load (the P1a spec's acceptance, finally run live).
• The frontdoor suite + the engine suites green (delta vs the baselines); drift gate PASS.

Prior art — #1806/PR #1807 (P1a, merged to the lane); #1803's port list (P1's frontdoor half); the #1264 lineage the ring replaced.

Source — Part of #1803 (P1b). Builds on the lane's merged P1a.