fix(gateway): /queue is now a true FIFO — each invocation gets its own turn

[teknium1]

Summary

Repeated /queue commands now each produce a full agent turn, in FIFO order, with no merging. Previously the second /queue silently overwrote the first because the handler wrote directly into the adapter's single-slot _pending_messages dict — CLI users got stacking (real queue.Queue) but every gateway platform dropped everything but the last item.

Changes

• gateway/run.py: GatewayRunner grows a _queued_events: Dict[str, List[MessageEvent]] overflow buffer. /queue uses the adapter's next-up slot when free, otherwise appends to overflow; drain promotes the overflow head into the slot after each consumed turn so the recursion sees it.
• _handle_reset_command (/new, /reset) clears the overflow for that session. /model and other mid-session operations preserve the queue.
• _handle_status_command (/status) reports Queued follow-ups: N when non-zero.
• /queue ack now surfaces depth once it exceeds 1 (Queued for the next turn. (3 queued)).
• Extracted _enqueue_fifo, _promote_queued_event, _queue_depth helpers. All use getattr(self, '_queued_events', None) fallback so existing tests using object.__new__(GatewayRunner) (pitfall Add support for Atropos Agentic RL environments (requires branch tool… #17) keep working.

Validation

	Before	After
/queue A; /queue B; /queue C on gateway	only C runs	A, B, C each get own turn
Text merging across /queue items	possible via merge_pending_message_event	impossible — separate dict slots
/new after queueing	overflow leaked into new conversation	cleared
/status output	no queue visibility	shows depth when non-zero
Targeted tests	n/a	50 pass (queue, drain race, hygiene, status, model reset, steer)

Added 5 new tests to tests/gateway/test_queue_consumption.py covering FIFO enqueue, drain-site promotion across 3 consecutive turns, interrupt-follow-up staging, depth counting, and no-merge invariant. Removed the stale test_multiple_queues_last_one_wins test which documented the old broken behavior.

CLI behavior is unchanged — self._pending_input = queue.Queue() already did this correctly.

[ether-btc]

Excellence Contribution — /queue True FIFO

What it does
Each /queue invocation now gets its own distinct agent turn, processed in strict FIFO order. Previously, the gateway wrote queued events directly into the adapter's single-slot _pending_messages dict — subsequent /queue commands overwrote prior ones. The CLI already used queue.Queue() and handled this correctly; all gateway platforms did not.

Why it matters
Queue starvation was a silent correctness bug. /queue A; /queue B; /queue C on any gateway platform (Discord, Slack, Telegram, etc.) would execute only C. A and B were silently discarded with no error, no warning. Any workflow using /queue to stage sequential operations — batch document processing, multi-step analysis — had operations silently disappear.

Technical approach
GatewayRunner gains a _queued_events: Dict[str, List[MessageEvent]] overflow buffer. When /queue fires and the adapter slot is occupied, the event appends to overflow. On each drain completion, the overflow head is promoted into the slot in order. /new clears the overflow; /status shows Queued follow-ups: N when N > 0.

Behavior	Before	After
/queue A; /queue B; /queue C on gateway	Only C runs	A, B, C each get their own turn
/new after queueing	Overflow leaked into new conversation	Overflow cleared
/status	No queue visibility	Shows Queued follow-ups: N
CLI behavior	Correct	Unchanged

Implications

• Sequential batch workflows on all gateway platforms now work correctly
• /queue graduates from "unreliable on gateway" to a first-class workflow primitive
• The Queued follow-ups visibility in /status makes queue state transparent — eliminating the silent-drop failure mode