docs(networking): fix round — argus review + audit E1/E2

observability.md: "Why two tiers" claimed the harness currently writes an
upstream token into the agent's own claude settings; that path was removed
with the direct-export mode it served (nix/agent-modules/otel.nix:56-63).
Reworded as the hypothetical the paragraph is actually making.

gateway.md: restored the agent-trust pointer to
/run/hive-ca/trust-bundle.pem in "Cert prompts" (hive-ca-trust.nix:41),
dropped by the earlier rewrite. Corrected the SPA-fallback section: only
the chat.<swarm> vhost uses the Accept-header map
(hive-matrix.nix:693-696); per-agent split mode uses file-existence
try_files (gateway_nginx.rs:93-134), not the same mechanism.

Refs #3902

docs/scheduler/observability.md

@@ -59,10 +59,11 @@ the control plane, so degraded telemetry isn't degraded operation.
 ### Why two tiers
 
 **The hive tier isn't optional.** Exporting straight to `endpoint` would mean
-every agent needs the credential — and the harness delivers that token into
-the agent's own `~/.claude/settings.json`, a file the agent can read. `0600`
-protects it from other containers, not from the agent itself. An option that
-could select the direct path would reopen that hole.
+every agent needs the credential — and the only place to hand it to an agent
+container is somewhere the agent itself can read, its own claude settings
+among them. `0600` protects a secret from other containers, not from the
+agent it belongs to. An option that could select that path would reopen the
+hole.
 
 **The tiers stay separate on one box.** An all-local hive is a statement about
 _where_ processes run, not about the shape of the deployment. A boundary that