feat(browser): stream bounded live session updates (#26)

docs/integration/live-session-updates.md

@@ -0,0 +1,114 @@
+# Live session-list updates
+
+Tracking: #26
+
+Live updates are an optional acceleration for an open server browser. The
+bounded `GET /v1/sessions` snapshot remains the source of truth, and join
+authorization still revalidates current capacity, presence, policy, and
+compatibility. A displayed player count is advisory, never an admission promise.
+
+## Snapshot, stream, reset
+
+Every `BrowseSessionsResponse` includes `streamCursor` in addition to its normal
+pagination cursor. Connect to `GET /v1/sessions/stream` with the same game,
+environment, protocol, optional region, and `excludeFull` filter. Send the most
+recent stream cursor as `Last-Event-ID`.
+
+| SSE event | Contract kind | UI action |
+| --- | --- | --- |
+| `session_upsert` | `sessionUpsert` | Add or replace the complete public projection by listing ID. |
+| `session_remove` | `sessionRemove` | Remove the listing ID. |
+| `reset` | `reset` | Discard local state, fetch a fresh snapshot, then reconnect with its cursor. |
+| `keepalive` | `keepalive` | Preserve the cursor and connection; do not change UI state. |
+
+Each SSE `id` equals the opaque cursor inside its JSON event. Cursors are signed,
+short-lived, monotonically ordered, and bound to the complete filter. A missing,
+expired, corrupted, foreign, future, or replay-gapped cursor produces `reset`
+instead of a potentially incomplete view. Do not parse or retain it as a stable
+identifier.
+
+Updates cover creation after fresh UDP presence, public-field/capacity changes,
+presence staleness and recovery, lease expiry, deregistration, operator or
+principal revocation, and visibility/region/protocol changes. Events contain the
+same bounded public `SessionListing` as snapshots. They never contain raw peer
+endpoints, lease tokens, punch capabilities, tickets, publisher subjects, or
+internal store identifiers.
+
+## SDK and polling fallback
+
+```csharp
+BrowseSessionsRequest filter = new()
+{
+    GameId = new("space-game"),
+    EnvironmentId = new("production"),
+    ProtocolVersion = 7,
+    RegionId = new("eu-central"),
+    ExcludeFull = true,
+};
+RendezvousClientResult<BrowseSessionsResponse> snapshot =
+    await browser.BrowseAsync(filter, cancellationToken);
+
+await foreach (RendezvousClientResult<SessionStreamEvent> update in
+    browser.StreamAsync(filter, snapshot.Value!.StreamCursor, cancellationToken))
+{
+    if (!update.IsSuccess)
+    {
+        // Switch to bounded polling with jittered backoff.
+        break;
+    }
+    // Apply upsert/remove by listing ID. On reset, discard and browse again.
+}
+```
+
+Cancellation or enumerator disposal closes the response and releases the server
+subscription. A normal connection-duration close is a reconnect signal: use the
+last applied event cursor. Repeated failures, unsupported platform HTTP stacks,
+and restrictive proxies fall back to snapshots with exponential jittered
+backoff, a capped interval, and `Retry-After`. Never open parallel streams to
+compensate for a slow UI.
+
+## TestClient
+
+```bash
+dotnet run --project src/FinalFactory.Rendezvous.TestClient \
+  --configuration Release --no-build -- \
+  watch --service https://rendezvous.example.invalid/ \
+  --game space-game --environment production --region eu-central --protocol 7 \
+  --run-seconds 60 --json
+```
+
+`watch.snapshot`, `watch.session-upsert`, `watch.session-remove`,
+`watch.keepalive`, and `watch.reconnect` are stable diagnostics. Add
+`--exercise-reset --script` to corrupt the snapshot cursor deliberately and
+verify a typed reset plus snapshot refresh. Use `--exercise-reconnect --script`
+while producing one update to close the first stream deliberately, reconnect
+from its prior cursor, and verify that the same ordered event is replayed.
+Polished list diffing, selection retention, animation, and accessibility remain
+in each game.
+
+## Bounds and slow consumers
+
+The v1 journal retains at most 4,096 public-only changes. It admits at most 256
+subscribers total and 64 per tenant, reads at most 128 changes per batch,
+waits a configurable 50 milliseconds after a live change and coalesces the
+resulting batch to the final change per listing, sends a keepalive every 15
+seconds, and closes a connection after five minutes. A consumer behind the
+replay window receives `reset`; it never acquires an unbounded queue.
+
+Normal optional-work concurrency and per-source/tenant rate controls apply for
+the stream lifetime. Exhaustion returns typed HTTP `429` before streaming.
+Shutdown cancels streams; reconnect only after readiness returns and expect a
+reset after a single-active restart because listings and replay are ephemeral.
+
+## Reverse proxy
+
+- Disable response buffering (`X-Accel-Buffering: no` is also emitted),
+  compression, transformation, and caching for `text/event-stream`.
+- Preserve `Last-Event-ID`; set upstream/read timeouts above the 15-second
+  keepalive and around six minutes for the five-minute connection ceiling.
+- Flush events promptly and use HTTP/2 only when streaming semantics survive.
+- Preserve the source-IP trust boundary and abuse controls; do not add a bypass.
+
+Verify the deployed proxy with an idle keepalive, update, reconnect, invalid
+cursor reset, slow reader, and graceful shutdown. An in-process pass does not
+prove that a production proxy is non-buffering.

docs/integration/test-client.md

@@ -146,7 +146,14 @@ Never use an unbounded sleep to orchestrate processes. Wait for versioned events
 such as `host.ready` and apply a deadline. Useful success events are
 `host.registered`, `host.ready`, `host.direct-traffic`, `host.deregistered`,
 `browse.completed`, `browse.session`, `join.connected`, `join.direct-traffic`,
-and `join.outcome-report`.
+`join.outcome-report`, `watch.snapshot`, `watch.session-upsert`,
+`watch.session-remove`, `watch.reset`, `watch.reconnect`, and `watch.complete`.
+
+For a bounded live-directory diagnostic, use `watch --run-seconds 60`. Add
+`--exercise-reset --script` to prove fail-closed cursor recovery, or
+`--exercise-reconnect --script` while changing one listing to prove ordered
+`Last-Event-ID` replay after a deliberate disconnect. The full event and proxy
+contract is in [live session-list updates](live-session-updates.md).
 
 | Exit | Meaning |
 | ---: | --- |