feat(memory): embedders vectoriels réels HTTP + ONNX derrière features (LOT C1)

Remplace les StubEmbedder pour les stratégies localServer/api/localOnnx par de
vrais moteurs, chacun derrière une feature cargo off-by-default — la posture
fondatrice « rien d'imposé, zéro dépendance » (défaut none → rappel naïf) reste
byte-for-byte inchangée.

C1a (feature vector-http, reqwest rustls optional):
- HttpEmbedder couvrant localServer (Ollama/llama.cpp) et api (OpenAI/Voyage…),
  payload OpenAI-compatible /v1/embeddings, ordre restauré par index, bearer
  token lu via env var (jamais en clair), timeout client 30s.
- detect_ollama() pour la détection de l'existant (C3).

C1b (feature vector-onnx, fastembed v5 optional):
- OnnxEmbedder en-process (e5-small, dim 384), init paresseuse + spawn_blocking,
  cache modèle sous <app_data>/embedders/onnx — aucun download au build ni au
  first-run, uniquement à la demande au 1er embed.
- Catalogue RECOMMENDED_ONNX_MODELS + ONNX_CACHE_SUBDIR + onnx_model_is_cached
  exposés (sans feature) pour la config (C2) et la popup (C3).

embedder_from_profile(profile, onnx_cache_dir) dispatche feature-gated ; sans la
feature, retombe sur StubEmbedder (Unsupported) → fallback naïf via
AdaptiveMemoryRecall. Composition root (build_memory_recall) propage le cache dir.

Tests: 10 HTTP + 6 ONNX (dont 2 #[ignore] download réel) + 26 vectoriels, verts
en défaut, --features vector-http et --features vector-onnx.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

crates/infrastructure/tests/vector_recall.rs

@@ -26,8 +26,9 @@ use domain::ports::{
 use domain::profile::{EmbedderProfile, EmbedderStrategy};
 use domain::project::ProjectPath;
 use infrastructure::{
-    embedder_from_profile, should_use_vector, AdaptiveMemoryRecall, FsMemoryStore, HashEmbedder,
-    NaiveMemoryRecall, StubEmbedder, VectorMemoryRecall,
+    embedder_from_profile, onnx_model_is_cached, should_use_vector, AdaptiveMemoryRecall,
+    FsMemoryStore, HashEmbedder, NaiveMemoryRecall, StubEmbedder, VectorMemoryRecall,
+    ONNX_CACHE_SUBDIR, RECOMMENDED_ONNX_MODELS,
 };
 
 // ---------------------------------------------------------------------------
@@ -255,26 +256,48 @@ async fn stub_embedder_returns_unsupported() {
 
 #[tokio::test]
 async fn embedder_from_profile_maps_each_strategy() {
-    // none ⇒ no embedder (recall stays naïve).
-    assert!(embedder_from_profile(&EmbedderProfile::none()).is_none());
+    // The ONNX cache dir is irrelevant to every strategy exercised here (none of
+    // these short-circuiting cases loads a model), so a throwaway path is fine.
+    let onnx_cache = std::path::Path::new("/unused-onnx-cache");
 
-    // localOnnx / localServer / api ⇒ a StubEmbedder (Unsupported), dimension kept.
-    for strategy in [
-        EmbedderStrategy::LocalOnnx,
-        EmbedderStrategy::LocalServer,
-        EmbedderStrategy::Api,
-    ] {
-        let profile =
-            EmbedderProfile::new("e", "E", strategy, None, None, None, 24).unwrap();
-        let embedder = embedder_from_profile(&profile)
-            .unwrap_or_else(|| panic!("{strategy:?} must yield an embedder"));
-        assert_eq!(embedder.dimension(), 24);
+    // none ⇒ no embedder (recall stays naïve).
+    assert!(embedder_from_profile(&EmbedderProfile::none(), onnx_cache).is_none());
+
+    // localOnnx without the `vector-onnx` feature maps to a StubEmbedder
+    // (Unsupported); with the feature it is the real OnnxEmbedder. Dimension is
+    // preserved either way (the real engine validates dimension only at embed time).
+    let onnx =
+        EmbedderProfile::new("e", "E", EmbedderStrategy::LocalOnnx, None, None, None, 24).unwrap();
+    let embedder =
+        embedder_from_profile(&onnx, onnx_cache).expect("localOnnx must yield an embedder");
+    assert_eq!(embedder.dimension(), 24);
+    #[cfg(not(feature = "vector-onnx"))]
+    {
         let err = embedder.embed(&["t".to_string()]).await.unwrap_err();
         assert!(
             matches!(err, EmbedderError::Unsupported(_)),
-            "{strategy:?} stub must be Unsupported, got {err:?}"
+            "localOnnx stub must be Unsupported without vector-onnx, got {err:?}"
         );
     }
+
+    // localServer / api: a stub *without* the `vector-http` feature (default,
+    // dependency-free build), a real HttpEmbedder *with* it. Either way an embedder
+    // is produced with the declared dimension.
+    for strategy in [EmbedderStrategy::LocalServer, EmbedderStrategy::Api] {
+        let profile = EmbedderProfile::new("e", "E", strategy, None, None, None, 24).unwrap();
+        let embedder = embedder_from_profile(&profile, onnx_cache)
+            .unwrap_or_else(|| panic!("{strategy:?} must yield an embedder"));
+        assert_eq!(embedder.dimension(), 24);
+
+        #[cfg(not(feature = "vector-http"))]
+        {
+            let err = embedder.embed(&["t".to_string()]).await.unwrap_err();
+            assert!(
+                matches!(err, EmbedderError::Unsupported(_)),
+                "{strategy:?} stub must be Unsupported without vector-http, got {err:?}"
+            );
+        }
+    }
 }
 
 // ===========================================================================
@@ -597,6 +620,93 @@ async fn adaptive_falls_back_to_naive_when_embedder_fails() {
     assert!(!got.is_empty(), "the degraded result still returns entries");
 }
 
+// ===========================================================================
+// 7. Always-available ONNX data/inspection helpers (compiled WITHOUT the
+//    `vector-onnx` feature — they are pure data + pure-FS, no `fastembed`).
+// ===========================================================================
+
+/// A unique, self-cleaning scratch dir under the OS temp dir (project convention;
+/// see tests/project_store.rs), used to exercise `onnx_model_is_cached` on real FS.
+struct TempDir(std::path::PathBuf);
+impl TempDir {
+    fn new() -> Self {
+        let p = std::env::temp_dir().join(format!("idea-onnx-cached-{}", uuid::Uuid::new_v4()));
+        std::fs::create_dir_all(&p).unwrap();
+        Self(p)
+    }
+    fn path(&self) -> &std::path::Path {
+        &self.0
+    }
+}
+impl Drop for TempDir {
+    fn drop(&mut self) {
+        let _ = std::fs::remove_dir_all(&self.0);
+    }
+}
+
+#[test]
+fn recommended_onnx_models_advertise_e5_small_384() {
+    assert_eq!(RECOMMENDED_ONNX_MODELS.len(), 1, "exactly one curated model");
+    let m = RECOMMENDED_ONNX_MODELS[0];
+    assert_eq!(m.id, "multilingual-e5-small");
+    assert_eq!(m.dimension, 384);
+    assert!(m.recommended, "the single curated model is the recommended default");
+}
+
+#[test]
+fn onnx_cache_subdir_is_stable() {
+    assert_eq!(ONNX_CACHE_SUBDIR, "embedders/onnx");
+}
+
+#[test]
+fn onnx_model_is_cached_false_on_empty_dir() {
+    // A fresh, empty cache dir ⇒ nothing cached.
+    let tmp = TempDir::new();
+    assert!(
+        !onnx_model_is_cached(tmp.path(), "multilingual-e5-small"),
+        "an empty cache dir reports no cached model"
+    );
+}
+
+#[test]
+fn onnx_model_is_cached_false_when_dir_missing() {
+    // A path that does not exist must report `false`, never panic.
+    let missing = std::path::Path::new("/idea-onnx-cache-definitely-missing-xyz");
+    assert!(!onnx_model_is_cached(missing, "multilingual-e5-small"));
+}
+
+#[test]
+fn onnx_model_is_cached_true_when_model_subdir_present_and_nonempty() {
+    // The heuristic looks for a non-empty subdirectory whose (lowercased) name
+    // *contains* the model token (slashes/underscores normalised to `-`), mirroring
+    // hf-hub's `models--<org>--<name>` layout.
+    let tmp = TempDir::new();
+    let model_dir = tmp.path().join("models--Qdrant--multilingual-e5-small-onnx");
+    std::fs::create_dir_all(&model_dir).unwrap();
+    // The subdir must be non-empty to count as "cached".
+    std::fs::write(model_dir.join("model.onnx"), b"not really a model").unwrap();
+
+    assert!(
+        onnx_model_is_cached(tmp.path(), "multilingual-e5-small"),
+        "a non-empty matching subdir ⇒ cached"
+    );
+}
+
+#[test]
+fn onnx_model_is_cached_false_when_matching_subdir_is_empty() {
+    // A matching but *empty* subdir does not count as a present cached model.
+    let tmp = TempDir::new();
+    std::fs::create_dir_all(tmp.path().join("models--Qdrant--multilingual-e5-small-onnx")).unwrap();
+    assert!(
+        !onnx_model_is_cached(tmp.path(), "multilingual-e5-small"),
+        "an empty matching subdir is not a cached model"
+    );
+}
+
+// ===========================================================================
+// 8. AdaptiveMemoryRecall — the étage-1/étage-2 switch (cont.).
+// ===========================================================================
+
 #[tokio::test]
 async fn adaptive_empty_memory_and_zero_budget_are_empty() {
     // Liskov common contract: empty list, never an error, both paths.