feat: add multimodal search and granular performance diagnostics

crates/searchi-semantic/tests/gemma_cpu.rs

@@ -1,6 +1,49 @@
 use anyhow::Result;
 use searchi_semantic::{Device, GemmaEmbedder, GEMMA_DIMENSIONS};
 
+#[test]
+#[ignore = "Requires explicitly provisioned Google model and CPU worker"]
+fn google_cpu_debug_query_reports_actual_worker_stages_without_changing_vectors() -> Result<()> {
+    let model = std::env::var_os("SEARCHI_GEMMA_MODEL_DIR").expect("Set SEARCHI_GEMMA_MODEL_DIR");
+    let worker = std::env::var_os("SEARCHI_GEMMA_WORKER").expect("Set SEARCHI_GEMMA_WORKER");
+    let mut embedder = GemmaEmbedder::load_with_device(model, worker, 1, Device::Cpu)?;
+    let query = "supplier contract cancellation policy";
+    let debug = embedder.embed_query_debug(query)?;
+    let diagnostics = embedder.query_diagnostics();
+    for name in [
+        "request_encode_us",
+        "request_write_us",
+        "worker_wait_and_response_decode_us",
+        "response_validate_us",
+        "worker_tokenization_us",
+        "worker_batch_prepare_us",
+        "worker_inference_us",
+        "worker_pool_normalize_us",
+        "worker_token_count",
+    ] {
+        assert!(
+            diagnostics.contains_key(name),
+            "Missing actual stage {name}: {diagnostics:?}"
+        );
+    }
+    assert!(diagnostics["worker_inference_us"] > 0);
+    assert!(diagnostics["worker_token_count"] > 0);
+    assert!(!serde_json::to_string(&diagnostics)?.contains(query));
+    let normal = embedder.embed_query(query)?;
+    assert!(embedder.query_diagnostics().is_empty());
+    let cosine: f64 = debug
+        .iter()
+        .zip(&normal)
+        .map(|(a, b)| f64::from(*a) * f64::from(*b))
+        .sum();
+    assert!(
+        cosine >= 0.99999,
+        "Debug tracing changed embeddings: {cosine}"
+    );
+    println!("CPU query diagnostics: {diagnostics:?}; debug/normal cosine {cosine:.8}");
+    Ok(())
+}
+
 #[test]
 #[ignore = "Requires explicitly provisioned Google model and CPU worker"]
 fn google_cpu_real_embeddings_and_verified_tail_windows() -> Result<()> {