Resolve adapters by what is mounted, and say why each kind is not loaded

runtime.py resolved adapters at 7 call sites and only ever looked in the
adapter library, so an adapter attached to a cartridge (cartridge.py:230
copies it into <cartridge>/adapters/{kind}/) was never loaded by anything.
Resolution now lives in adapters.py: it reads the mounted cartridge set,
prefers a mounted cartridge's copy over the library, and answers every one
of the six kinds with a state the facts support — loaded, available,
deferred (base_model mismatch or missing weights, both models named in the
reason), missing — plus a shadowed row for each adapter that lost the
choice. A mounted cartridge is the scenario selection; neither manifest
carries a scenario field, so none was invented.

Two bugs the new tests caught: the preference sort lost reverse=True, so a
library adapter beat a mounted cartridge's copy, and the load branch had
never been executed by a test. Stub weights now prove peft is asked to
attach, that the returned model is put in eval mode, and that the endpoint
reports loaded until unload clears it.

Faces: GET /api/runtime/adapters, `ely-eye adapters`, and a log line naming
the reason whenever a generation runs without an adapter.

Gates: 222 passed; ruff clean; mypy cold cache on the 7 touched files gives
the same 2 errors as HEAD's runtime.py, and pyright gives the same 2 as
HEAD — both are qwen_vl_utils typing artifacts (process_vision_info returns
2 or 3 values depending on return_video_kwargs; the product takes 2).

CHECKLIST.md

@@ -219,10 +219,19 @@
       重跑 `bench.asr_check`（新进程）、重跑门禁、逐项核对 stamp 链与 18 条 CHECKLIST 断言，
       并确认旧数字（1.7 秒 / 3.55 / 61.6 / 31 段 / 210.5 / 51×）零残留。三轮都是同一个 reviewer，
       没有换人放水。
-- [ ] **p1-3** adapter 加载
+- [x] **p1-3** adapter 加载
-  - [ ] p1-3.1 按卡带/场景解析 6 个 adapter，不再只有 `runtime.py:317`
+  - [x] p1-3.1 按卡带/场景解析 6 个 adapter，不再只有 `runtime.py:317`
-  - [ ] p1-3.2 加载结果可观测：日志 + `GET /api/runtime/adapters`
+  - [x] p1-3.2 加载结果可观测：日志 + `GET /api/runtime/adapters`
   - 验收：状态端点列每个 adapter 的 loaded/deferred + 原因；桩权重测试证明加载分支执行
+  - `backend/src/ely_eye/adapters.py` 是唯一的 adapter 解析处。`runtime.py` 原先在 7 处调 `resolve_runtime_adapter`，现在统一走 `_generation_adapter_fields()`；`peft` 只在 `adapters.load_into_model` 里懒加载，`RuntimeAdapter` 与 `resolve_runtime_adapter` 已删除。
+  - 「场景」= 挂载的卡带：`AdapterManifest` 和 `CartridgeManifest` 都没有 scenario 字段（读出来的事实），所以解析按 `cartridge_store.mounted_ids` 选，报告写明来自哪个卡带，没有新造字段。
+  - 修掉的真 bug：`CartridgeService.attach_adapter`（`cartridge.py:230`）把 adapter 拷进 `<卡带>/adapters/{kind}/`，但解析从来不看卡带目录，挂上去的东西运行时根本不加载。现在挂载卡带里的副本优先于库里的同名 adapter。
+  - 测试写出来才发现的第二个 bug：`ready.sort(key=_preference)` 丢了原来的 `reverse=True`，库里的 adapter 反而赢了挂载副本。
+  - 状态只由事实支撑：loaded（有 LoadRecord）/ available（manifest 与 `adapter_model.safetensors` 都在）/ deferred（base_model 不匹配或缺权重文件，原因里写两个模型名）/ missing（磁盘上没有该 kind）。落选者单独一行 `shadowed=True`。
+  - 验收命令：`pytest backend/tests/test_adapters.py -q` → 10 passed。`test_stub_weights_prove_the_load_branch_runs` 用桩权重断言 `attached == [(ANY, 路径)]`、`model.evalled is True`、LoadRecord 存在、端点该行 `state == "loaded"`；`test_unload_clears_the_loaded_state` 断言 unload 后回到 available。
+  - 新文件 `backend/src/ely_eye/runtime_schemas.py`（30 行）：`schemas.py` 已 498 行，离 500 行红线 2 行，报告类型不放进去。
+  - 门禁：`pytest -q` 222 passed；ruff 干净；mypy 冷缓存查 7 个改动文件 → 2 errors，与 HEAD 的 `runtime.py` 逐条相同（`qwen_vl_utils` 无 stubs、`model_class` 重赋值），净新增 0；pyright 同 7 文件 → 2 errors，HEAD 也是 2 条，都是 `process_vision_info` 的双返回（`vision_process.py:532-534` 由 `return_video_kwargs` 决定返回 2 或 3 元组，产品用默认值取 2 个，运行时正确）。
+  - CLI 面：`ely-eye adapters` 打印同一份报告。
 - [ ] **p1-4** 工程地基
   - [ ] p1-4.1 pytest 从零到有（后端）
   - [ ] p1-4.2 go test 从零到有（MCP）

backend/src/ely_eye/adapters.py

@@ -0,0 +1,229 @@
+"""One place decides which adapter loads, and says out loud why the others do not."""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+from pydantic import ValidationError
+
+from . import cartridge_store
+from .config import Settings
+from .db import Database
+from .runtime_schemas import AdapterChoice, AdapterLoadReport, AdapterOrigin, AdapterState
+from .schemas import AdapterManifest, TrainingAdapterKind
+
+logger = logging.getLogger("ely_eye.adapters")
+
+MANIFEST_FILE = "adapter_manifest.json"
+WEIGHTS_FILE = "adapter_model.safetensors"
+
+
+@dataclass(frozen=True)
+class AdapterSource:
+    """One adapter found on disk, plus where it was found."""
+
+    path: Path
+    manifest: AdapterManifest
+    origin: AdapterOrigin
+    cartridge_id: str | None
+
+
+@dataclass(frozen=True)
+class LoadRecord:
+    """An adapter this process actually attached to a live model."""
+
+    kind: TrainingAdapterKind
+    adapter_id: str
+    path: str
+    sha256: str
+    loaded_at: str
+
+
+_LOADED: dict[str, LoadRecord] = {}
+
+
+def read_manifest(path: Path) -> AdapterManifest:
+    try:
+        text = path.read_text(encoding="utf-8")
+    except OSError as error:
+        raise ValueError(f"{path}: the adapter manifest cannot be read ({error})") from error
+    try:
+        return AdapterManifest.model_validate_json(text)
+    except ValidationError as error:
+        raise ValueError(f"{path}: the adapter manifest is not a valid AdapterManifest ({error})") from error
+
+
+def cartridge_roots(database: Database) -> dict[str, Path]:
+    roots: dict[str, Path] = {}
+    for row in database.list_cartridges():
+        roots[str(row["cartridge_id"])] = Path(str(row["root_path"]))
+    return roots
+
+
+def library_sources(settings: Settings) -> list[AdapterSource]:
+    """Adapters in the library directory, for callers that have no database open."""
+    if settings.adapter_dir is not None:
+        pinned = settings.adapter_dir.expanduser()
+        if not (pinned / MANIFEST_FILE).exists():
+            logger.warning("adapter_dir %s holds no %s", pinned, MANIFEST_FILE)
+            return []
+        manifest = read_manifest(pinned / MANIFEST_FILE)
+        return [
+            AdapterSource(
+                path=pinned,
+                manifest=manifest,
+                origin="library",
+                cartridge_id=manifest.cartridge_id,
+            )
+        ]
+    sources: list[AdapterSource] = []
+    for manifest_path in sorted(settings.adapters_dir.glob(f"*/{MANIFEST_FILE}")):
+        manifest = read_manifest(manifest_path)
+        sources.append(
+            AdapterSource(
+                path=manifest_path.parent,
+                manifest=manifest,
+                origin="library",
+                cartridge_id=manifest.cartridge_id,
+            )
+        )
+    return sources
+
+
+def discover(settings: Settings, database: Database) -> tuple[list[AdapterSource], list[str]]:
+    """Every adapter the runtime could use: the library plus the copies inside mounted cartridges."""
+    mounted = cartridge_store.mounted_ids(database)
+    roots = cartridge_roots(database)
+    sources = library_sources(settings)
+    for cartridge_id in mounted:
+        root = roots.get(cartridge_id)
+        if root is None:
+            logger.warning("cartridge %s is mounted but its directory is not registered", cartridge_id)
+            continue
+        for manifest_path in sorted((root / "adapters").glob(f"*/{MANIFEST_FILE}")):
+            sources.append(
+                AdapterSource(
+                    path=manifest_path.parent,
+                    manifest=read_manifest(manifest_path),
+                    origin="cartridge",
+                    cartridge_id=cartridge_id,
+                )
+            )
+    return sources, mounted
+
+
+def resolve(
+    sources: list[AdapterSource],
+    settings: Settings,
+    kind: TrainingAdapterKind,
+) -> tuple[AdapterSource | None, AdapterState, str]:
+    """Pick the adapter to load for one kind, or say which state it is stuck in and why."""
+    candidates = [source for source in sources if source.manifest.kind == kind]
+    if not candidates:
+        return None, "missing", f"no adapter of kind {kind.value} is on disk"
+    matching = [source for source in candidates if source.manifest.base_model == settings.model_id]
+    if not matching:
+        trained_on = ", ".join(sorted({source.manifest.base_model for source in candidates}))
+        return (
+            None,
+            "deferred",
+            f"every {kind.value} adapter is trained on {trained_on}; the runtime holds {settings.model_id}",
+        )
+    ready = [source for source in matching if (source.path / WEIGHTS_FILE).exists()]
+    if not ready:
+        return None, "deferred", f"the {kind.value} manifests are present but none carries {WEIGHTS_FILE}"
+    ready.sort(key=_preference, reverse=True)
+    chosen = ready[0]
+    if chosen.origin == "cartridge":
+        return chosen, "available", f"selected from mounted cartridge {chosen.cartridge_id}"
+    return chosen, "available", "selected from the adapter library"
+
+
+def resolve_library_adapter(
+    settings: Settings,
+    kind: TrainingAdapterKind = TrainingAdapterKind.hme_core_lora,
+) -> tuple[AdapterSource | None, AdapterState, str]:
+    """Resolution without a database: the library only, no mounted cartridges."""
+    return resolve(library_sources(settings), settings, kind)
+
+
+def _preference(source: AdapterSource) -> tuple[bool, float]:
+    weights = source.path / WEIGHTS_FILE
+    return (source.origin == "cartridge", weights.stat().st_mtime if weights.exists() else 0.0)
+
+
+def load_records() -> list[LoadRecord]:
+    return list(_LOADED.values())
+
+
+def clear_loads() -> None:
+    _LOADED.clear()
+
+
+def record_load(source: AdapterSource) -> LoadRecord:
+    record = LoadRecord(
+        kind=source.manifest.kind,
+        adapter_id=source.manifest.adapter_id,
+        path=str(source.path),
+        sha256=source.manifest.sha256,
+        loaded_at=datetime.now().astimezone().isoformat(),
+    )
+    _LOADED[record.kind.value] = record
+    logger.info("loaded adapter %s (%s) from %s", record.adapter_id, record.kind.value, record.path)
+    return record
+
+
+def load_into_model(model: Any, path: Path) -> Any:
+    """Attach one LoRA to a live model; the only place generation imports peft."""
+    from .training import bitsandbytes_gaudi_probe_guard
+
+    with bitsandbytes_gaudi_probe_guard():
+        from peft import PeftModel
+
+        return PeftModel.from_pretrained(model, str(path))
+
+
+def plan(settings: Settings, database: Database) -> AdapterLoadReport:
+    """One row per adapter kind, plus a row for every adapter that lost the choice."""
+    sources, mounted = discover(settings, database)
+    rows: list[AdapterChoice] = []
+    for kind in TrainingAdapterKind:
+        source, state, reason = resolve(sources, settings, kind)
+        record = next((item for item in load_records() if item.kind == kind), None)
+        if record is not None:
+            reason = f"attached to the live model from {record.path}"
+            state = "loaded"
+        rows.append(
+            AdapterChoice(
+                kind=kind,
+                state=state,
+                reason=reason,
+                path=str(source.path) if source else None,
+                adapter_id=source.manifest.adapter_id if source else None,
+                base_model=source.manifest.base_model if source else settings.model_id,
+                origin=source.origin if source else None,
+                cartridge_id=source.cartridge_id if source else None,
+            )
+        )
+    chosen_paths = {row.path for row in rows if row.path}
+    for source in sources:
+        if str(source.path) in chosen_paths:
+            continue
+        rows.append(
+            AdapterChoice(
+                kind=source.manifest.kind,
+                state="deferred",
+                reason="another adapter of this kind wins for the current mount",
+                path=str(source.path),
+                adapter_id=source.manifest.adapter_id,
+                base_model=source.manifest.base_model,
+                origin=source.origin,
+                cartridge_id=source.cartridge_id,
+                shadowed=True,
+            )
+        )
+    return AdapterLoadReport(model_id=settings.model_id, mounted=mounted, adapters=rows)

backend/src/ely_eye/api.py

@@ -10,6 +10,7 @@ from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 
 from . import __version__
+from .adapters import plan as adapter_plan
 from .api_cartridges import register as register_cartridge_routes
 from .api_holders import register as register_holder_routes
 from .benchmarks import (
@@ -41,6 +42,7 @@ from .proofs import HashHopEvaluator, PRDProofSuite
 from .research_modules import ResearchModuleService
 from .retrieval import RetrievalService
 from .runtime import QwenRuntime, RuntimeUnavailableError
+from .runtime_schemas import AdapterLoadReport
 from .schemas import (
     AdapterManifest,
     AdapterStatus,
@@ -277,6 +279,10 @@ def create_app() -> FastAPI:
     def runtime_profiles() -> RuntimeProfileReport:
         return runtime_profile_report()
 
+    @app.get("/api/runtime/adapters", response_model=AdapterLoadReport)
+    def runtime_adapters(settings: Annotated[Settings, Depends(settings_dep)]) -> AdapterLoadReport:
+        return adapter_plan(settings, Database(settings))
+
     @app.get("/api/research/modules")
     def research_modules(settings: Annotated[Settings, Depends(settings_dep)]):
         db = Database(settings)

backend/src/ely_eye/cli.py

@@ -6,6 +6,7 @@ from pathlib import Path
 import typer
 import uvicorn
 
+from .adapters import plan as adapter_plan
 from .api import create_app
 from .benchmarks import (
     benchmark_prediction_report,
@@ -475,5 +476,13 @@ def attach_adapter(cartridge_id: str, adapter_dir: Path) -> None:
     typer.echo(manifest.model_dump_json(indent=2))
 
 
+@app.command("adapters")
+def adapter_status() -> None:
+    """Which adapter the runtime would load, and why each other kind would not."""
+    settings = get_settings()
+    report = adapter_plan(settings, Database(settings))
+    typer.echo(json_dump(report.model_dump(mode="json")))
+
+
 if __name__ == "__main__":
     app()

backend/src/ely_eye/proofs/checks_runtime.py

@@ -9,6 +9,7 @@ from __future__ import annotations
 from pathlib import Path
 from typing import Any
 
+from .. import adapters
 from ..compiler import ContextCompiler
 from ..config import Settings
 from ..db import Database
@@ -17,7 +18,7 @@ from ..evaluation_schemas import (
 )
 from ..evidence_graph import EvidenceGraphService
 from ..profiles import runtime_profile_report
-from ..runtime import QwenRuntime, resolve_runtime_adapter
+from ..runtime import QwenRuntime
 from ..schemas import (
     RuntimeProfile,
     TrainingAdapterKind,
@@ -58,7 +59,7 @@ class RuntimeChecks:
             return deferred
         runtime = QwenRuntime(self.settings)
         status = runtime.status()
-        adapter = resolve_runtime_adapter(self.settings)
+        adapter, _adapter_state, _adapter_reason = adapters.resolve_library_adapter(self.settings)
         weight_path = adapter.path / "adapter_model.safetensors" if adapter else None
         payload = {
             "runtime_status": status.model_dump(mode="json"),
@@ -133,7 +134,7 @@ class RuntimeChecks:
             )
             payload["generation_question"] = generation_question
             status = runtime.status()
-            adapter = resolve_runtime_adapter(proof_settings)
+            adapter, _adapter_state, _adapter_reason = adapters.resolve_library_adapter(proof_settings)
             generation = runtime.generate(generation_question, context)
             answer, citations, citation_report = Verifier().verify_answer(generation.text, context)
             claim_id = EvidenceGraphService(proof_settings, self.db).record_answer(

backend/src/ely_eye/runtime.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import base64
 import json
+import logging
 import subprocess
 from dataclasses import dataclass
 from pathlib import Path
@@ -9,10 +10,14 @@ from typing import Any
 
 import httpx
 
+from . import adapters
 from .config import Settings, get_settings
-from .schemas import AdapterManifest, CompiledContext, RuntimeStatus, TrainingAdapterKind
+from .db import Database
+from .schemas import CompiledContext, RuntimeStatus, TrainingAdapterKind
 from .storage import ObjectStore
 
+logger = logging.getLogger("ely_eye.runtime")
+
 
 class RuntimeUnavailableError(RuntimeError):
     pass
@@ -34,12 +39,41 @@ class QwenRuntime:
     def __init__(self, settings: Settings | None = None) -> None:
         self.settings = settings or get_settings()
         self.objects = ObjectStore(self.settings)
+        self.database = Database(self.settings)
         self._processor: Any | None = None
         self._model: Any | None = None
 
+    def _adapter_choice(self) -> tuple[adapters.AdapterSource | None, str]:
+        """The adapter this runtime would attach, plus the reason if it would not attach one."""
+        sources, _ = adapters.discover(self.settings, self.database)
+        source, _state, reason = adapters.resolve(sources, self.settings, TrainingAdapterKind.hme_core_lora)
+        return source, reason
+
+    def _loaded_adapter(self) -> adapters.LoadRecord | None:
+        return next((item for item in adapters.load_records() if item.kind == TrainingAdapterKind.hme_core_lora), None)
+
+    def _generation_adapter_fields(self) -> dict[str, str | None]:
+        record = self._loaded_adapter()
+        if record is not None:
+            return {
+                "adapter_path": record.path,
+                "adapter_kind": record.kind.value,
+                "adapter_id": record.adapter_id,
+                "adapter_sha256": record.sha256,
+            }
+        source, _ = self._adapter_choice()
+        if source is None:
+            return {"adapter_path": None, "adapter_kind": None, "adapter_id": None, "adapter_sha256": None}
+        return {
+            "adapter_path": str(source.path),
+            "adapter_kind": source.manifest.kind.value,
+            "adapter_id": source.manifest.adapter_id,
+            "adapter_sha256": source.manifest.sha256,
+        }
+
     def status(self) -> RuntimeStatus:
         gpu = gpu_status()
-        adapter = resolve_runtime_adapter(self.settings)
+        fields = self._generation_adapter_fields()
         if self.settings.runtime_backend == "sglang":
             available, detail = self._sglang_status()
         else:
@@ -49,10 +83,10 @@ class QwenRuntime:
             model_id=self.settings.model_id,
             available=available,
             detail=detail,
-            adapter_path=str(adapter.path) if adapter else None,
+            adapter_path=fields["adapter_path"],
-            adapter_kind=adapter.manifest.kind if adapter else None,
+            adapter_kind=TrainingAdapterKind(fields["adapter_kind"]) if fields["adapter_kind"] else None,
-            adapter_id=adapter.manifest.adapter_id if adapter else None,
+            adapter_id=fields["adapter_id"],
-            adapter_sha256=adapter.manifest.sha256 if adapter else None,
+            adapter_sha256=fields["adapter_sha256"],
             gpu_name=gpu.get("name"),
             cuda=gpu.get("cuda"),
             vram_total_mb=gpu.get("total_mb"),
@@ -84,15 +118,17 @@ class QwenRuntime:
             import torch
 
             adapter_detail = ""
-            adapter = resolve_runtime_adapter(self.settings)
+            source, reason = self._adapter_choice()
-            if adapter is not None:
+            if source is not None:
-                adapter_detail = f" Adapter selected: {adapter.path}."
+                adapter_detail = f" Adapter selected: {source.path}."
             elif self.settings.adapter_dir is not None:
                 return False, f"Configured adapter path is missing: {self.settings.adapter_dir}"
+            if source is None:
+                adapter_detail = f" No adapter: {reason}."
             if torch.cuda.is_available():
                 return True, f"Transformers backend can use CUDA.{adapter_detail}"
             return True, f"Transformers backend is available on CPU.{adapter_detail}"
-        except Exception as exc:
+        except Exception as exc:  # noqa: BLE001  # a status probe reports why the backend is unusable
             return False, f"Transformers backend import failed: {exc}"
 
     def _generate_sglang(self, question: str, context: CompiledContext) -> RuntimeGeneration:
@@ -122,15 +158,11 @@ class QwenRuntime:
         except httpx.HTTPError as exc:
             raise RuntimeUnavailableError(f"SGLang generation failed: {exc}") from exc
         data = response.json()
-        adapter = resolve_runtime_adapter(self.settings)
         usage = data.get("usage") if isinstance(data, dict) else None
         return RuntimeGeneration(
             text=data["choices"][0]["message"]["content"],
             backend="sglang",
-            adapter_path=str(adapter.path) if adapter else None,
+            **self._generation_adapter_fields(),
-            adapter_kind=adapter.manifest.kind.value if adapter else None,
-            adapter_id=adapter.manifest.adapter_id if adapter else None,
-            adapter_sha256=adapter.manifest.sha256 if adapter else None,
             input_tokens=usage.get("prompt_tokens") if isinstance(usage, dict) else None,
             output_tokens=usage.get("completion_tokens") if isinstance(usage, dict) else None,
         )
@@ -161,15 +193,11 @@ class QwenRuntime:
         except httpx.HTTPError as exc:
             raise RuntimeUnavailableError(f"SGLang generation failed: {exc}") from exc
         data = response.json()
-        adapter = resolve_runtime_adapter(self.settings)
         usage = data.get("usage") if isinstance(data, dict) else None
         return RuntimeGeneration(
             text=data["choices"][0]["message"]["content"],
             backend="sglang",
-            adapter_path=str(adapter.path) if adapter else None,
+            **self._generation_adapter_fields(),
-            adapter_kind=adapter.manifest.kind.value if adapter else None,
-            adapter_id=adapter.manifest.adapter_id if adapter else None,
-            adapter_sha256=adapter.manifest.sha256 if adapter else None,
             input_tokens=usage.get("prompt_tokens") if isinstance(usage, dict) else None,
             output_tokens=usage.get("completion_tokens") if isinstance(usage, dict) else None,
         )
@@ -218,14 +246,10 @@ class QwenRuntime:
         input_length = int(inputs["input_ids"].shape[-1])
         generated = generated[:, input_length:]
         decoded = processor.batch_decode(generated, skip_special_tokens=True)[0]
-        adapter = resolve_runtime_adapter(self.settings)
         return RuntimeGeneration(
             text=decoded,
             backend="transformers",
-            adapter_path=str(adapter.path) if adapter else None,
+            **self._generation_adapter_fields(),
-            adapter_kind=adapter.manifest.kind.value if adapter else None,
-            adapter_id=adapter.manifest.adapter_id if adapter else None,
-            adapter_sha256=adapter.manifest.sha256 if adapter else None,
             input_tokens=input_length,
             output_tokens=int(generated.shape[-1]),
         )
@@ -273,14 +297,10 @@ class QwenRuntime:
         input_length = int(inputs["input_ids"].shape[-1])
         generated = generated[:, input_length:]
         decoded = processor.batch_decode(generated, skip_special_tokens=True)[0]
-        adapter = resolve_runtime_adapter(self.settings)
         return RuntimeGeneration(
             text=decoded,
             backend="transformers",
-            adapter_path=str(adapter.path) if adapter else None,
+            **self._generation_adapter_fields(),
-            adapter_kind=adapter.manifest.kind.value if adapter else None,
-            adapter_id=adapter.manifest.adapter_id if adapter else None,
-            adapter_sha256=adapter.manifest.sha256 if adapter else None,
             input_tokens=input_length,
             output_tokens=int(generated.shape[-1]),
         )
@@ -288,6 +308,19 @@ class QwenRuntime:
     def _load_transformers(self) -> tuple[Any, Any]:
         if self._model is not None and self._processor is not None:
             return self._model, self._processor
+        model, processor = self._load_base_model()
+        source, reason = self._adapter_choice()
+        if source is None:
+            logger.info("generating without an adapter: %s", reason)
+        else:
+            model = adapters.load_into_model(model, source.path)
+            adapters.record_load(source)
+        model.eval()
+        self._model, self._processor = model, processor
+        return model, processor
+
+    def _load_base_model(self) -> tuple[Any, Any]:
+        """Load the chat model; tests replace this so the adapter branch runs without a 9B download."""
         import torch
         from transformers import AutoProcessor
 
@@ -300,33 +333,25 @@ class QwenRuntime:
 
             model_cls = AutoModelForCausalLM
 
-        self._processor = AutoProcessor.from_pretrained(self.settings.model_id, trust_remote_code=True)
+        processor = AutoProcessor.from_pretrained(self.settings.model_id, trust_remote_code=True)
-        self._model = model_cls.from_pretrained(
+        model = model_cls.from_pretrained(
             self.settings.model_id,
             dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
             device_map="auto",
             trust_remote_code=True,
         )
-        adapter = resolve_runtime_adapter(self.settings)
+        return model, processor
-        if adapter is not None:
-            from .training import bitsandbytes_gaudi_probe_guard
-
-            with bitsandbytes_gaudi_probe_guard():
-                from peft import PeftModel
-
-                self._model = PeftModel.from_pretrained(self._model, adapter.path)
-        self._model.eval()
-        return self._model, self._processor
 
     def unload(self) -> None:
         self._model = None
         self._processor = None
+        adapters.clear_loads()
         try:
             import torch
 
             if torch.cuda.is_available():
                 torch.cuda.empty_cache()
-        except Exception:
+        except Exception:  # noqa: BLE001  # freeing the cache is best-effort, unloading must not raise
             return
 
     def _context_images(self, context: CompiledContext) -> list[Path]:
@@ -338,43 +363,6 @@ class QwenRuntime:
         return paths
 
 
-@dataclass(frozen=True)
-class RuntimeAdapter:
-    path: Path
-    manifest: AdapterManifest
-
-
-def resolve_runtime_adapter(
-    settings: Settings,
-    kind: TrainingAdapterKind = TrainingAdapterKind.hme_core_lora,
-) -> RuntimeAdapter | None:
-    if settings.adapter_dir is not None:
-        path = settings.adapter_dir.expanduser().resolve()
-        manifest_path = path / "adapter_manifest.json"
-        if path.exists() and manifest_path.exists():
-            return RuntimeAdapter(
-                path=path,
-                manifest=AdapterManifest.model_validate_json(manifest_path.read_text(encoding="utf-8")),
-            )
-        return None
-
-    candidates: list[RuntimeAdapter] = []
-    for manifest_path in settings.adapters_dir.glob("*/adapter_manifest.json"):
-        manifest = AdapterManifest.model_validate_json(manifest_path.read_text(encoding="utf-8"))
-        if manifest.kind == kind and manifest.base_model == settings.model_id:
-            candidates.append(RuntimeAdapter(path=manifest_path.parent.resolve(), manifest=manifest))
-    candidates.sort(
-        key=lambda item: (
-            item.manifest.cartridge_id is not None,
-            (item.path / "adapter_model.safetensors").stat().st_mtime
-            if (item.path / "adapter_model.safetensors").exists()
-            else 0.0,
-        ),
-        reverse=True,
-    )
-    return candidates[0] if candidates else None
-
-
 def build_prompt(question: str, context: CompiledContext) -> str:
     contract = "\n".join(f"- {item}" for item in context.plan.verifier_contract)
     return f"""You are Ely-Eye, a local multimodal context operating system.

backend/src/ely_eye/runtime_schemas.py

@@ -0,0 +1,30 @@
+"""Runtime observability shapes: what the model holds in memory right now."""
+
+from typing import Literal
+
+from pydantic import BaseModel
+
+from .schemas import TrainingAdapterKind
+
+AdapterState = Literal["loaded", "available", "deferred", "missing"]
+AdapterOrigin = Literal["library", "cartridge"]
+
+
+class AdapterChoice(BaseModel):
+    """One adapter kind, and the honest reason it is or is not attached to the model."""
+
+    kind: TrainingAdapterKind
+    state: AdapterState
+    reason: str
+    path: str | None = None
+    adapter_id: str | None = None
+    base_model: str | None = None
+    origin: AdapterOrigin | None = None
+    cartridge_id: str | None = None
+    shadowed: bool = False
+
+
+class AdapterLoadReport(BaseModel):
+    model_id: str
+    mounted: list[str]
+    adapters: list[AdapterChoice]

backend/tests/test_adapters.py

@@ -0,0 +1,227 @@
+"""Which adapter the runtime loads, and the reason every other kind does not."""
+
+from __future__ import annotations
+
+import json
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+from unittest import mock
+
+import pytest
+from fastapi.testclient import TestClient
+
+from ely_eye import adapters, api
+from ely_eye.cartridge_store import set_mounts
+from ely_eye.config import Settings
+from ely_eye.db import Database
+from ely_eye.runtime import QwenRuntime
+from ely_eye.schemas import CartridgeManifest, TrainingAdapterKind
+
+CORE = TrainingAdapterKind.hme_core_lora
+MODEL = "Qwen/Qwen3.5-9B"
+
+
+def _manifest(kind: TrainingAdapterKind, base_model: str, adapter_id: str) -> dict[str, Any]:
+    return {
+        "adapter_id": adapter_id,
+        "kind": kind.value,
+        "created_at": "2026-05-01T00:00:00+00:00",
+        "base_model": base_model,
+        "dataset_path": "PRD.md",
+        "dataset_sha256": "d" * 64,
+        "sample_count": 8,
+        "max_steps": 16,
+        "max_length": 512,
+        "rank": 8,
+        "alpha": 16,
+        "target_modules": ["q_proj"],
+        "trainable_params": 1_000,
+        "total_params": 10_000,
+        "optimizer": "adamw_8bit",
+        "learning_rate": 0.0002,
+        "gradient_accumulation_steps": 1,
+        "bf16": True,
+        "final_loss": 0.01,
+        "artifacts": {},
+        "sha256": "a" * 64,
+    }
+
+
+def _write_adapter(
+    root: Path,
+    kind: TrainingAdapterKind = CORE,
+    base_model: str = MODEL,
+    adapter_id: str = "adapter-1",
+    weights: bool = True,
+) -> Path:
+    root.mkdir(parents=True, exist_ok=True)
+    manifest = _manifest(kind, base_model, adapter_id)
+    (root / "adapter_manifest.json").write_text(json.dumps(manifest), encoding="utf-8")
+    if weights:
+        (root / "adapter_model.safetensors").write_bytes(b"stub weights")
+    return root
+
+
+def _settings(tmp_path: Path) -> Settings:
+    settings = Settings(home=tmp_path / "home", model_id=MODEL)
+    settings.ensure_dirs()
+    return settings
+
+
+def _cartridge(settings: Settings, database: Database, cartridge_id: str) -> Path:
+    root = settings.home / "cartridges" / cartridge_id
+    root.mkdir(parents=True, exist_ok=True)
+    now = datetime.now(UTC)
+    manifest = CartridgeManifest(
+        cartridge_id=cartridge_id,
+        name="home",
+        created_at=now,
+        model_base=settings.model_id,
+        source_count=0,
+        atom_count=0,
+        token_equivalent=0,
+        artifacts={},
+        dna=None,
+    )
+    database.write_cartridge(cartridge_id, "home", root, manifest.model_dump_json(), None, 0, now)
+    set_mounts(database, [cartridge_id], now)
+    return root
+
+
+class _FakeModel:
+    def __init__(self) -> None:
+        self.evalled = False
+
+    def eval(self) -> None:
+        self.evalled = True
+
+
+@pytest.fixture(autouse=True)
+def _no_stale_loads():
+    adapters.clear_loads()
+    yield
+    adapters.clear_loads()
+
+
+def test_a_library_adapter_is_selected_for_its_kind(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    chosen = _write_adapter(settings.adapters_dir / "core")
+    source, state, reason = adapters.resolve_library_adapter(settings)
+    assert source is not None and source.path == chosen
+    assert state == "available"
+    assert reason == "selected from the adapter library"
+
+
+def test_an_adapter_on_another_base_model_is_deferred_naming_both_models(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    _write_adapter(settings.adapters_dir / "router", kind=TrainingAdapterKind.hme_router, base_model="Qwen/Qwen3.5-0.8B")
+    source, state, reason = adapters.resolve_library_adapter(settings, TrainingAdapterKind.hme_router)
+    assert source is None
+    assert state == "deferred"
+    assert "trained on Qwen/Qwen3.5-0.8B" in reason and f"runtime holds {MODEL}" in reason
+
+
+def test_a_manifest_without_weights_is_deferred_naming_the_missing_file(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    _write_adapter(settings.adapters_dir / "core", weights=False)
+    source, state, reason = adapters.resolve_library_adapter(settings)
+    assert source is None and state == "deferred"
+    assert "adapter_model.safetensors" in reason
+
+
+def test_the_mounted_cartridge_copy_wins_over_the_library(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    database = Database(settings)
+    _write_adapter(settings.adapters_dir / "core", adapter_id="library-copy")
+    cartridge_root = _cartridge(settings, database, "cart-home")
+    cartridge_copy = _write_adapter(cartridge_root / "adapters" / "hme_core_lora", adapter_id="cartridge-copy")
+    sources, mounted = adapters.discover(settings, database)
+    assert mounted == ["cart-home"]
+    source, state, reason = adapters.resolve(sources, settings, CORE)
+    assert source is not None and source.path == cartridge_copy
+    assert state == "available" and "mounted cartridge cart-home" in reason
+    report = adapters.plan(settings, database)
+    row = next(item for item in report.adapters if item.kind == CORE)
+    assert row.adapter_id == "cartridge-copy" and row.cartridge_id == "cart-home"
+
+
+def test_the_endpoint_lists_every_kind_with_a_reason(tmp_path: Path, monkeypatch) -> None:
+    settings = _settings(tmp_path)
+    _write_adapter(settings.adapters_dir / "core")
+    monkeypatch.setattr(api, "get_settings", lambda: settings)
+    client = TestClient(api.create_app())
+    body = client.get("/api/runtime/adapters").json()
+    assert body["model_id"] == MODEL and body["mounted"] == []
+    by_kind = {row["kind"]: row for row in body["adapters"]}
+    assert set(by_kind) == {kind.value for kind in TrainingAdapterKind}
+    assert by_kind[CORE.value]["state"] == "available"
+    for kind in TrainingAdapterKind:
+        if kind == CORE:
+            continue
+        assert by_kind[kind.value]["state"] == "missing"
+        assert by_kind[kind.value]["reason"], f"{kind.value} carries no reason"
+
+
+def test_stub_weights_prove_the_load_branch_runs(tmp_path: Path, monkeypatch) -> None:
+    settings = _settings(tmp_path)
+    chosen = _write_adapter(settings.adapters_dir / "core")
+    attached: list[tuple[Any, Path]] = []
+
+    def fake_attach(model: Any, path: Path) -> _FakeModel:
+        attached.append((model, path))
+        return _FakeModel()
+
+    monkeypatch.setattr(adapters, "load_into_model", fake_attach)
+    monkeypatch.setattr(QwenRuntime, "_load_base_model", lambda self: (_FakeModel(), object()))
+    runtime = QwenRuntime(settings)
+    model, _processor = runtime._load_transformers()
+    assert attached == [(mock.ANY, chosen)], "peft was never asked to attach the stub adapter"
+    assert model.evalled is True
+    assert [record.adapter_id for record in adapters.load_records()] == ["adapter-1"]
+    row = next(item for item in adapters.plan(settings, Database(settings)).adapters if item.kind == CORE)
+    assert row.state == "loaded" and "attached to the live model" in row.reason
+
+
+def test_unload_clears_the_loaded_state(tmp_path: Path, monkeypatch) -> None:
+    settings = _settings(tmp_path)
+    _write_adapter(settings.adapters_dir / "core")
+    monkeypatch.setattr(adapters, "load_into_model", lambda model, path: _FakeModel())
+    monkeypatch.setattr(QwenRuntime, "_load_base_model", lambda self: (_FakeModel(), object()))
+    runtime = QwenRuntime(settings)
+    runtime._load_transformers()
+    assert adapters.load_records(), "nothing was recorded as loaded"
+    runtime.unload()
+    assert adapters.load_records() == []
+    row = next(item for item in adapters.plan(settings, Database(settings)).adapters if item.kind == CORE)
+    assert row.state == "available"
+
+
+def test_the_loser_of_the_choice_is_reported_as_shadowed(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    _write_adapter(settings.adapters_dir / "first", adapter_id="written_first")
+    _write_adapter(settings.adapters_dir / "second", adapter_id="written_second")
+    report = adapters.plan(settings, Database(settings))
+    losers = [row for row in report.adapters if row.shadowed]
+    assert [row.adapter_id for row in losers] == ["written_first"]
+    assert "another adapter of this kind wins" in losers[0].reason
+    winner = next(row for row in report.adapters if row.adapter_id == "written_second")
+    assert winner.shadowed is False and winner.state == "available"
+
+
+def test_an_unreadable_manifest_names_the_file(tmp_path: Path) -> None:
+    settings = _settings(tmp_path)
+    broken = settings.adapters_dir / "broken"
+    broken.mkdir(parents=True)
+    (broken / "adapter_manifest.json").write_text("{not json", encoding="utf-8")
+    with pytest.raises(ValueError, match="the adapter manifest is not a valid AdapterManifest"):
+        adapters.library_sources(settings)
+
+
+def test_a_pinned_adapter_dir_without_a_manifest_says_the_path_is_missing(tmp_path: Path) -> None:
+    settings = Settings(home=tmp_path / "home", model_id=MODEL, adapter_dir=tmp_path / "empty")
+    settings.ensure_dirs()
+    assert adapters.library_sources(settings) == []
+    available, detail = QwenRuntime(settings)._transformers_status()
+    assert available is False
+    assert detail == f"Configured adapter path is missing: {tmp_path / 'empty'}"