Web UI WS state machine: outbox + seq + ack + keepalive

[seamus-brady]

Summary

Replaces the per-node monotonic seq (diagnostic-only) with a per-client outbox keyed on stable client_id. Every server-pushed frame is appended to that client's outbox before being shipped, the client acks by seq, the server prunes acked frames. On reconnect the client opens with ?since=N; the server replays the gap from the outbox or falls back to a full session_history rebuild when the gap exceeds the buffer.

Adds WS keepalive — server emits Ping every 25s, client replies Pong. Idle proxies and OS-level NAT can no longer silently drop the connection during long agent cycles.

Three reported symptoms this fixes

Symptom	Root cause	Fix
Long-running cycle, no live notification, refresh shows the reply	WS dropped silently (no keepalive); reply fires into dead notify_subject	Keepalive prevents the dropout; if it happens anyway, reconnect-with-since replays the missed frames
Typed user message disappears	user_message_ack dropped from inFlightUserMessages BEFORE cog persisted; race with reconnect-mid-cycle wiping the DOM	Ack is purely a UI hint; in-flight tracking stays until renderSessionHistory observes the message in the server's authoritative view
No live agent_progress / thinking / tool updates during work	Same as #1 — notifications fire into a dead WS	Same fix — keepalive + outbox replay

What's in the box

• src/web/outbox.gleam — pure ring buffer with seq, ack, replay_since (UpToDate / Replay / TooOld with boundary handling), age-prune, size cap. 12 unit tests.
• src/web/client_registry.gleam — OTP actor wrapping per-client_id Outboxes. Survives WS process death so reconnects under the same id replay correctly. Hourly janitor drops idle clients.
• src/web/gui.gleam — WsState gains client_id, registry, keepalive_subject. New ws_send helper routes every message through client_registry.append. ~25 direct mist.send_text_frame call sites converted. ?since=N parsed from URL; ws_on_init either spawns the history query (fresh / TooOld) or fires ReplayReady (gap fits). New Ack(seq) and Pong client-message handlers. KeepaliveTick selector arm emits Ping and re-arms.
• src/web/protocol.gleam — encode_server_message_with_seq takes explicit seq; encode_server_message_body returns the seq-less body for the outbox; splice_seq re-emits a stored body with its original seq during replay. Old global monotonic_seq FFI removed; encode_server_message kept as a back-compat alias that emits seq=0.
• src/web/html.gleam — both ws_connect_js (chat + admin) and the mobile page's connect() track lastSeenSeq in sessionStorage, build wsUrl with &since=N, send periodic {type:"ack", seq:N} every 5s, reply pong on {type:"ping"}. user_message_ack handler no longer drops in-flight entries.

What is preserved

The full UI surface — uploads (POST /upload + attachment refs in user_message), all admin tabs (Narrative, Log, Scheduler, Cycles, Planner, D' Safety, D' Config, Comms, Affect, Skills, Memory, Documents), question/answer flow, history browsing, search, approve/reject — all unchanged. Every server message type retains its existing JSON shape; only the seq field's semantics changed (per-node → per-client).

Test plan

• gleam build clean
• gleam format clean
• gleam test — 2190 passing (gained 12 outbox tests + 11 from concurrent main merges)
• Rebased cleanly on latest origin/main
• gleam run — confirm boot, observe [outbox] debug entries on first WS connect
• Live exercise the three reported symptoms:
  • Send a slow query, watch agent_progress arrive live (no refresh needed)
  • Reconnect mid-cycle (kill the network briefly), verify response still lands
  • Type a message during a long cycle, verify the bubble persists across the eventual reply

🤖 Generated with Claude Code