Handshake Budget: One RTT Against the Cover Sheet

Status: Note, REV A. Not an RFC. It annotates the cover sheet. The 100 ms ceiling from exchange tick to phone screen has always been drawn as if the socket already exists. On a handset, that is often false. Setup is a line item, and today it is unbudgeted.

The cold path we actually ship

The flight edge speaks WebSocket. A cold client pays, in order:

1. TCP handshake.
2. TLS handshake.
3. HTTP upgrade to the WebSocket.

That is multiple round trips before the first application byte. The WebTransport overview and the FOSDEM 2026 walkthrough of the same protocol put a number on the alternative: HTTP/3 and QUIC merge the transport and cryptographic handshakes into one RTT. A resumed session can send application data in 0-RTT. Those are protocol facts. They are not yet parts on this manifest.

Do the arithmetic in the open. A phone with a 40 ms round trip, which is an ordinary LTE or congested Wi-Fi figure and not a lab loopback, spends 120 ms on three RTTs and has already missed the cover sheet. No tick has been normalized. No shard has been consulted. The handset is still negotiating the right to listen. Cross-region adds more. The public dashboard already shows that region, not the normalize step, dominates p99. Setup is region. It belongs on the same strip.

What one RTT does not buy

Cutting setup to a single round trip does not make quotes faster once the session is warm. A warm WebSocket already holds the TCP session. The steady-state budget is still ingest, normalize, shard, fan-out, edge, paint. One RTT matters at:

• Cold start of the reference client.
• Radio change, when the phone leaves Wi-Fi and the TCP session dies.
• Silent NAT timeout, when the middlebox drops a flow that heartbeated too slowly.
• Region failover, when the edge the handset pinned is no longer the edge.

QUIC connection migration is the part that would keep a session across a radio change without a new handshake. It is real in the transport. It is not flight here. The reference client reconnects and requests a snapshot. Marking migration FLIGHT because a slide said "connection IDs" would be a documentation defect. This note records it as HOLD, same as raw UDP on the manifest.

0-RTT is a sharper hold. Resumption replays the bytes you sent early. An idempotent "latest quote for this symbol" can survive replay. A subscribe that mutates the interest index cannot. Replay would mint a second interest entry, which RFC-001 forbids by the one-owner rule. Until the comment ledger has a replay rule, 0-RTT stays off the flight binary. One RTT for a cold authenticated session is the saving we can take without lying about replay.

Where the ceiling should be redrawn

Propose four setup rows under the existing 100 ms ceiling, so the sheet stops pretending the socket is free:

Row	Part	Cold budget	Warm budget
Session	RT-300	1 RTT on the QUIC path, else measured upgrade	0, session cached
Snapshot	RT-310	one reliable exchange after session	only on gap or resume
Fan-out	RT-210	unchanged, 12 ms	unchanged
Paint	RT-600	inside the handset, not the bus	inside the handset

If the cold session row exceeds one RTT, the cover sheet is already blown and the rest of the ledger is fiction for that launch. Publish cold and warm separately. A single p99 that averages a long-lived desktop socket with a phone that just left a tunnel is how the strip stays green while the reference client stutters at the open.

Heartbeats stay a few seconds, not 30. A quiet flow that the NAT forgets is indistinguishable from a shard failure, and the reconnect pays the cold budget again. Two missed heartbeats drop the interest entry. That rule is already on RFC-001. This note only adds that the cost of being wrong is a full handshake, not a repaint.

See the cold start, not the average

The production consumer is Stock Picks. Force-quit it, open it on a phone that is not on the same LAN as the edge, and time the gap before the first quote. That gap is the unbudgeted line. If it is longer than the cover sheet, no amount of shard tuning inside the bus will move the number the handset actually feels. Comments on this note should argue with the table, not with a vibe about QUIC.