fix(simplex): make polling recovery visible and bounded

[Zoffka]

Summary

This PR tightens the SimpleX gateway fallback path for cases where daemon push events are missing or stale but /tail still exposes the message history.

It does not claim to fix every SimpleX daemon transport stall. During live testing, the local simplex-chat daemon itself could get contact-transport stale (sndSent stuck, no fresh inbound in daemon history) and recovered after restarting only the daemon. This PR focuses on the Hermes-side failures that are reproducible and covered by tests.

Changes

• Add a polling fallback for missed SimpleX chat items using /tail 50.
• Bound poll commands with short connect/read timeouts plus a hard wall-clock timeout so a stuck ephemeral websocket cannot stall the poll loop.
• Persist composite seen keys (direct:<contactId>:<itemId>, group:<groupId>:<itemId>) so restarts do not replay old mail.
• Do not seed unread inbound text as already seen; unread user messages remain actionable after restart/reconnect.
• Dispatch each polled batch in daemon order while keeping the poll loop free for the next tick.
• Mark polled inbound items read with /_read chat items ... so SimpleX users get a visible read indicator.
• Resolve outbound numeric Hermes chat IDs to SimpleX display-name targets before sending.
• Add a delayed SimpleX processing notice for slow turns because SimpleX has no typing indicator.
• Add an optional per-adapter stale active-session cap, used by SimpleX, so a fresh polled DM cannot sit behind an old stuck turn indefinitely.

Verification

Local tests:

47 passed   tests/gateway/test_simplex_plugin.py
288 passed  tests/gateway/test_simplex_plugin.py tests/gateway/test_platform_base.py tests/gateway/test_api_server.py

The broader gateway slice was run with live API-server CORS environment overrides unset so default-CORS tests used their expected defaults.

Live SimpleX checks on a local gateway:

• Outbound SimpleX marker delivered and acknowledged as sndRcvd.
• Fresh DM test ingested from SimpleX and replied to.
• Fresh DM wow ingested from SimpleX and replied to.
• Fresh group message test here ingested from SimpleX and replied to.
• Restart recovery test: an unread inbound item that predated gateway restart was left unseeded, dispatched on the first poll tick, marked read, and replied to.

Observed daemon-level caveat:

• In one failure mode, Hermes was still polling but the local simplex-chat daemon did not show the fresh contact message in /tail, and outbound messages stayed sndSent rather than sndRcvd. Restarting only simplex-chat.service restored delivery. That daemon transport-health issue is intentionally not hidden by this PR and should be handled separately if needed.

[muddlebee]

why?

[teknium1]

Thanks for the detailed SimpleX recovery work. I verified that the underlying gap still exists on current main: SimplexAdapter.connect() starts only _ws_listener() and _health_monitor() (plugins/platforms/simplex/adapter.py:235-238), and inbound delivery still depends on pushed newChatItems / newChatItem events (plugins/platforms/simplex/adapter.py:421-436).

Problems

• The PR’s poll-thread dispatch does not actually hop back to the gateway loop. PR head plugins/platforms/simplex/adapter.py:319-342 runs polling inside a dedicated thread event loop, but _dispatch_polled_items() then sees that loop via asyncio.get_running_loop() and creates the dispatch task on it (plugins/platforms/simplex/adapter.py:791-800). That means _handle_new_chat_item() / handle_message() can run on the poll-thread loop; the intended run_coroutine_threadsafe() gateway-loop branch (plugins/platforms/simplex/adapter.py:802-809) is bypassed from the poller.
• A salvage needs to preserve current main’s newer SimpleX handler behavior: correlated command responses are handled at plugins/platforms/simplex/adapter.py:370-390, and file descriptor/completion events at plugins/platforms/simplex/adapter.py:406-467.

Suggested changes

• Schedule polled dispatch on self._gateway_loop whenever the current loop is not the gateway loop; keep direct create_task() only for gateway-loop/test calls.
• Port the polling pieces into the current adapter without replacing the newer event/file handling paths.

This is an automated hermes-sweeper review.