feat(overlays): make llama.cpp router HF-cache scan opt-in

The router unconditionally scans the HF cache (server-models.cpp
load_from_cache) and auto-registers every downloaded model under a
forced repo:tag id, creating untuned twins of preset-defined models and
blocking free naming. The scan is just a listing convenience -
auto-download on first use goes through the child's --hf-repo and is
unaffected.

Patch gates it behind LLAMA_ROUTER_SCAN_CACHE (off by default) so
--models-preset is the single source of truth and models can be named
freely.

Assisted-by: pi (opus-4.8)

overlays/default.nix

@@ -63,6 +63,12 @@ in {
       })
     ];
 
+    # Make the llama.cpp router's HF-cache scan opt-in (LLAMA_ROUTER_SCAN_CACHE)
+    # so --models-preset is the single source of truth and models can be named
+    # freely without untuned repo:tag twins. Patch the base so the -vulkan and
+    # -rocm variants (llama-cpp.override) inherit it.
+    llama-cpp = addPatches prev.llama-cpp [./llama-cpp-optional-cache-scan.patch];
+
     wl-clipboard = addPatches prev.wl-clipboard [./wl-clipboard-secrets.diff];
 
     pass = addPatches prev.pass [./pass-wlclipboard-secret.diff];

overlays/llama-cpp-optional-cache-scan.patch

@@ -0,0 +1,29 @@
+Make the router's HF-cache scan opt-in.
+
+By default `llama-server` in router mode scans the HF cache and auto-registers
+every downloaded model under a forced `repo:tag` id, producing untuned
+duplicates of models already defined via --models-preset and preventing free
+naming. The scan is purely a listing convenience and is independent of
+auto-download on first use (that goes through the child process's --hf-repo).
+
+Gate it behind LLAMA_ROUTER_SCAN_CACHE so presets are the single source of
+truth by default; set the env var to restore upstream behaviour.
+
+--- a/tools/server/server-models.cpp
++++ b/tools/server/server-models.cpp
+@@ -280,8 +280,13 @@
+ 
+ void server_models::load_models() {
+     // Phase 1: load presets from all sources — pure I/O, no lock needed
+-    // 1. cached models
+-    common_presets cached_models = ctx_preset.load_from_cache();
++    // 1. cached models (opt-in). Scanning the HF cache auto-registers untuned
++    //    twins of models already defined as presets and forces their id to
++    //    repo:tag. Off by default; set LLAMA_ROUTER_SCAN_CACHE to restore it.
++    common_presets cached_models;
++    if (std::getenv("LLAMA_ROUTER_SCAN_CACHE")) {
++        cached_models = ctx_preset.load_from_cache();
++    }
+     SRV_INF("Loaded %zu cached model presets\n", cached_models.size());
+     // 2. local models from --models-dir
+     common_presets local_models;