feat(focus): single-channel focus mode with held backlog, autoreply, and expiry

changelog.d/feat-focus-mode.added.md

@@ -0,0 +1,38 @@
+- `focus` tool (opt-in via `FrameworkConfig.focus.enabled`): the primary
+  resident narrows attention to one channel for a bounded time. Every other
+  channel and DM is held — stored with `metadata.focusHeld` and excluded from
+  the resident's view, never woken on, and dropped by the wake gate before
+  any debounce policy can batch it. Addressed messages elsewhere get one
+  host-authored automatic reply per (channel, author) per epoch. Focus ends
+  on `mode: "end"`, at its deadline (durable across restarts), or when an
+  expired epoch is found at boot; ending delivers a per-channel
+  `<focus-backlog>` of the newest `backlogCap` held messages and points at
+  the resident's actual history/backscroll tools for the rest. `mode:
+  "check"` reports held counts (and peeks one channel) without ending;
+  `enter` while focused re-targets, delivering the new channel's backlog.
+- While focused, plain speech lands in the focus channel: it outranks the
+  trigger channel and the global most-recent-inbound fallback for every
+  turn, a successful `enter` moves the current turn's prose there too
+  (announced in the tool result), and held inbound neither retargets the
+  fallback locus nor qualifies for the mid-turn addressed re-pin.
+- Review hardening: channel identity is `(serverId, channelId)` throughout
+  (hold, gate, `enter` with an optional `serverId`, server-qualified
+  autoreply publish); the gate keeps each queued event's metadata so the
+  focus-entry purge, delivery-time re-check and live evaluation derive a
+  push event's channel identically; a held message never reaches a live
+  turn through the mid-turn deferred injection; held messages still deferred
+  behind the current turn are part of the unfocus backlog, and a deferred
+  focus dump makes the deferred queue durable so a crash cannot lose it; the
+  autoreply speaks only in channels the resident is open in and in DMs, and
+  a failed publish releases the per-author allowance; channels under
+  tune-out are left to the subconscious (and cannot be focused); the tool is
+  not offered under per-channel conversation routing; a persisted epoch
+  still ends at boot after `focus.enabled` is turned off; configured maxima
+  are floored/capped integers; media blocks are named in the dump.
+  Speech under focus is routed server-qualified (a same-id channel on
+  another server cannot receive it); automatic replies that published are
+  recorded in the lifecycle log, so once-per-author survives a restart and
+  a failed publish is retried, and a publish settling after its epoch ended
+  touches nothing; a held message awaiting its dump counts as unread for
+  event coalescing (an edit replaces it in place); wakes already queued for
+  a channel are dropped when that channel becomes held.