stuck-agent detector counts emitter-only unread as stuck - exemption tests the recipient, never the sender

STUCK_EXEMPT_RE = /^(?:scrp-|applog-)/i at hub.mjs:10593 is tested against the RECIPIENT (`agent`), never the SENDER. Its own comment states the intent as 'scrp-*/applog-* are dead-letter toasts with no reply path, so an undrained tail on one is not a fault' - that rationale describes the message rail, and the rail half is not implemented. Result: any agent whose entire unread tail is emitter traffic (scrp-* announces, llmmsg-srv-hub alerts) is counted stuck and re-alerted every SILENCE_REALERT_S forever, though nothing is wrong and there is nothing to act on. Measured 2026-09-11 13:41 UTC, one minute, ten simultaneous [stuck-agent] alerts: nw-venus-cc, nw-whey-cc, nw-lezama-cc, bin-venus-cc, coder-llmmsgsrv-cc, coder-pluto-cc, coder-terra-cc, monitor-context-cc, coder-enamel-cc, db-enamel-cc. nw-venus-cc was BUSY with a live spinner at the time (scrp-monitor-agent-idle.sh, state=BUSY prompt=''), i.e. a false positive on a working agent. The tails that do NOT self-clear are emitter-only by construction: mirty-cc 1 unread from an emitter aged 1336min, monitor-llmmsgsrv-cc 1 from an emitter aged 371min, nw-lezama-cc 2 of 2 from emitters aged 40min. Those three alert forever. Cost: each alert is ~300 tokens to the routed recipient plus the PM copy, and the flood lands on the nw agents that are themselves the alert target - it inflates the burn it is meant to surface (nw-venus-cc was flagged at 40006 tok/min vs a 2518 median in the same window). Scope: make the stuck predicate ignore unread whose sender matches the emitter pattern - i.e. count only unread that has a reply path. An agent holding ONLY emitter traffic is not stuck. Do not widen STUCK_EXEMPT_RE to more recipient names and do not add a disable flag; the recipient-side exemption is correct for its own reason (write-only mailboxes) and stays. Keep the journal line either way. Watch: a tail that is emitter-only must still clear stuck state, not merely skip the alert, or the reset branch stops running for it (same REV3 reasoning already in the file). And do not make the negative case silent - if every unread is filtered out, log it, so 'no alert' is distinguishable from 'nobody examined'. Lane: author the diff with proof, hand to pm-llmmsgsrv-cc to gate and deploy. Never applied by the coder.