Use measured query-only context across desktop CLI and MCP

AGENTS.md

@@ -51,6 +51,13 @@ readiness without completed release gates and recorded evidence.
   thread; yield between document encodes. Trace admission separately from native
   inference. This does not bound decoder, extraction or ANN CPU work. Do not
   claim preemption of an already executing media encode on a one-thread budget.
+- Foreground desktop, CLI and MCP sessions use the normal worker's query role
+  with a 512-token context; document/media indexing retains 2048 tokens. Preserve
+  role on CPU restart and reject indexing operations before projector work.
+  `media_supported` describes compiled capability, not a loaded projector or
+  permission to index in a query session. Legacy workers need an explicit update,
+  never a silent larger-context fallback. Keep private commitment, resident memory,
+  model verification and cached startup measurements distinct.
 - Treat native filesystem events as bounded hints; retain work across busy scans
   and verify periodically. Use targeted, cancellable reconciliation for events.
   Media hints must advance durable change revisions even with identical stats;