feat: add multimodal search and granular performance diagnostics

crates/searchi-semantic/tests/fixtures/README.md

@@ -0,0 +1,14 @@
+# Pinned Google media reference fixtures
+
+The `digup-photo.jpg`, `digup-screen.png`, `digup-speech.wav`, and
+`digup-media-reference.json` files are copied unchanged from
+`ARahim3/DigUp`, commit `6faa29e4c315a3a591de1b41bd78751bdf39588d`,
+`Tests/LlamaRuntimeTests/Reference/`. They and the existing text reference
+fixture are retained under the adjacent `LICENSE-DigUp.txt` MIT notice.
+
+The source test uses image budgets 280 (screen), 140 (photo), 70 (frame),
+and a mono 16 kHz speech recording. The native CPU test compares actual
+768-dimensional Google vectors with the pinned llama.cpp b11461 reference
+at cosine >=0.999. This is runtime parity evidence, not a production media
+relevance corpus. Independent images, audio, and video require separate
+quality evaluation.