Studio: track the loaded GGUF filename so native companion resolution reproduces the load identity

studio/backend/core/inference/sd_cpp_backend.py

@@ -220,6 +220,11 @@ class _SdState:
     mode: str = "server"
     # Token kept so LoRA adapters selected at generate time can be fetched from the Hub.
     hf_token: Optional[str] = None
+    # The single-file GGUF basename this load committed. Kept so companion resolution
+    # (sd_cpp_text_encoders_for) reproduces the load identity -- some variants pick their
+    # encoder by filename (FLUX.2-klein-9B -> Qwen3-8B), and a local *klein-9B*.gguf carries
+    # that keyword only in the basename, not the repo id.
+    gguf_filename: Optional[str] = None
 
 
 def _memory_policy(memory_mode: Optional[str], cpu_offload: bool) -> str:
@@ -570,6 +575,7 @@ class SdCppDiffusionBackend:
                     server = server,
                     mode = mode,
                     hf_token = hf_token,
+                    gguf_filename = gguf_filename,
                 )
                 with self._lock:
                     if self._load_token != _load_token:
@@ -697,10 +703,15 @@ class SdCppDiffusionBackend:
             repos = [state.repo_id, state.base_repo]
             if fam.sd_cpp_vae:
                 repos.append(fam.sd_cpp_vae[0])
-            # Same per-variant encoder selection as _asset_specs, keyed on the loaded repo id, so the
-            # cache-deletion guard protects the encoder repo this load actually downloaded (the 9B
-            # variant's Qwen3-8B, not the 4B default).
-            repos.extend(terepo for terepo, _f, _k in sd_cpp_text_encoders_for(fam, state.repo_id))
+            # Same per-variant encoder selection as _asset_specs, keyed on the loaded repo id AND
+            # GGUF filename, so the cache-deletion guard protects the encoder repo this load actually
+            # downloaded (the 9B variant's Qwen3-8B, not the 4B default) -- a local *klein-9B*.gguf
+            # carries that keyword only in the basename, so dropping the filename would fall back to
+            # the 4B default and protect the wrong repo.
+            repos.extend(
+                terepo
+                for terepo, _f, _k in sd_cpp_text_encoders_for(fam, state.repo_id, state.gguf_filename)
+            )
             return tuple(dict.fromkeys(r for r in repos if r))
 
     # ── Generate ───────────────────────────────────────────────────────────

studio/backend/tests/test_sd_cpp_backend.py

@@ -99,6 +99,31 @@ def test_loaded_repo_ids_includes_native_companions():
     assert b.loaded_repo_ids() == ()
 
 
+def test_loaded_repo_ids_tracks_variant_encoder_by_gguf_filename():
+    # FLUX.2-klein-9B pairs with Qwen3-8B, and a local *klein-9B*.gguf carries that keyword only in
+    # the basename (not the repo id). loaded_repo_ids() must reproduce the committed load identity
+    # (repo id + GGUF filename), else it falls back to the 4B default encoder and the cache-deletion
+    # guard protects the wrong repo -- leaving the 8B encoder this load actually downloaded deletable
+    # while one-shot sd-cli still re-reads it every generation.
+    b = SdCppDiffusionBackend(engine = _FakeEngine())
+    fam = detect_family("flux.2-klein")
+    b._state = bk._SdState(
+        repo_id = "local/my-klein-checkpoints",  # no variant keyword; it lives in the filename
+        base_repo = fam.base_repo,
+        family = fam,
+        device = "cpu",
+        files = SdCppModelFiles(diffusion_model = "/m/FLUX.2-klein-9B-Q4_K_M.gguf"),
+        vae_format = fam.sd_cpp_vae_format,
+        sampling_method = fam.sd_cpp_sampling_method,
+        flow_shift = fam.sd_cpp_flow_shift,
+        mode = "oneshot",
+        gguf_filename = "FLUX.2-klein-9B-Q4_K_M.gguf",
+    )
+    ids = set(b.loaded_repo_ids())
+    assert "Comfy-Org/vae-text-encorder-for-flux-klein-9b" in ids  # the 8B encoder this load pulled
+    assert "Comfy-Org/z_image_turbo" not in ids  # the 4B default must not be protected instead
+
+
 class _FakeServer:
     """Stands in for SdCppServer: records the spawn + one img_gen per whole batch."""