fix(audit): bump quinn-proto + ignore rmcp DNS-rebinding advisory

Two new RUSTSEC advisories landed between 2026-06-18 and 2026-06-30
that started failing every PR's `cargo audit` step:

- RUSTSEC-2026-0185 (quinn-proto 0.11.14): remote memory exhaustion via
  unbounded out-of-order stream reassembly. Patched in 0.11.15 (semver-
  compatible). `cargo update -p quinn-proto`, no API change.

- RUSTSEC-2026-0189 (rmcp 0.16.0): DNS rebinding in Streamable HTTP
  server transport due to missing Host-header validation. Patched in
  rmcp >= 1.4.0, which is a major API jump from our pin (rmcp shipped
  0.x -> 1.x -> 2.x in three months and the migration touches every
  tool handler + the auth middleware just landed in #92).

  Added to ignore with justification: our MCP server is exposed at a
  public hostname behind orca's TLS-terminating ingress with per-tenant
  bearer auth. The attack model (browser DNS-rebinding into a victim's
  localhost MCP server) doesn't apply to a public-hostname deployment.
  Defense-in-depth Host-header validation remains worthwhile, tracked
  as a multi-hour M7.3 follow-up to migrate rmcp 0.16 -> 2.x.

Unblocks #96 and #94.

.cargo/audit.toml

@@ -7,4 +7,17 @@ ignore = [
     # not a realistic attack surface here. Revisit when mongodb bumps hickory.
     "RUSTSEC-2026-0118", # NSEC3 loop, no fix available upstream
     "RUSTSEC-2026-0119", # O(n²) name compression, fixed in hickory-proto >=0.26.1
+
+    # rmcp 0.16.0 — DNS rebinding in Streamable HTTP server transport (missing
+    # Host header validation). Patched in rmcp >= 1.4.0, which is a major API
+    # version jump from our pin; rmcp shipped 0.x → 1.x → 2.x in three months
+    # and the migration touches every tool handler + the auth middleware we
+    # just landed in #92. Threat model in our deployment: the MCP server is
+    # exposed at a public hostname (comp-mcp-dev.meghsakha.com) behind orca's
+    # TLS-terminating ingress with per-tenant bearer auth — the attack model
+    # (browser DNS-rebinding into localhost MCP server) doesn't directly apply.
+    # Defense-in-depth Host-header check is still a worthwhile follow-up.
+    # FOLLOW-UP: bump rmcp to 2.x in a dedicated PR (M7.3 follow-up, sized
+    # multi-hour due to API surface change).
+    "RUSTSEC-2026-0189",
 ]