feat(fabric): coalesce a split agent turn into ONE message (deterministic)

OpenClaw delivers an agent turn whose blocks are text -> thinking/tool
-> text via multiple inbound deliver() calls (a non-text block is a
delivery boundary), so one turn became N Fabric messages.

Fix: buffer deliver() segments per channel (src/coalesce.ts) and flush
them as ONE postMessage at a deterministic boundary — the finally after
dispatchInboundReplyWithBase() resolves, which provably runs only after
every deliver() of the turn (verified: deliver,deliver -> dispatch
returned -> flush). No hooks, no timers, no idle guessing. The
agent_end hook was rejected: it fires BEFORE deliver(). gateway_stop
flushes any leftover; a long safety timeout is a leak-guard only.
channels.fabric.coalesce=false restores raw per-segment posting.

Verified on local openclaw + Fabric with a fake text/thinking/text
model: single trigger -> exactly one merged message.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

openclaw.plugin.json

@@ -44,6 +44,10 @@
             "minLength": 1,
             "description": "Shared secret that must equal the guild's FABRIC_BACKEND_GUILD_COMMANDS_SYNC_KEY. Required to register the slash-command catalog (Guild C-2). Read it from the guild via: docker exec fabric-backend-guild node dist/cli/print-commands-sync-key.js"
           },
+          "coalesce": {
+            "type": "boolean",
+            "description": "Merge a split agent turn (text → thinking/tool → text) into ONE Fabric message. Flushed deterministically on the agent_end hook. Default true; false = raw per-segment posting."
+          },
           "dmSecurity": { "type": "string" },
           "dmPolicy": { "type": "string" },
           "enabled": { "type": "boolean" },