fix(lmstudio): force encoding_format=float to avoid SDK base64 decode

The OpenAI SDK (≥6.x, since openai-node#1312) auto-injects
`encoding_format: "base64"` into every embeddings request when the
caller doesn't specify one, then unconditionally decodes the response
with `toFloat32Array(embedding as unknown as string)`.

LM Studio's Local Server ignores `encoding_format` and always returns a
JSON array of floats. The SDK then runs `Buffer.from(<array>, 'base64')`
— Node.js silently drops the encoding parameter for array inputs and
clamps each float (<1.0) to uint8 0, producing a 4096-byte zero buffer
that gets reinterpreted as a 1024-element Float32Array of zeros.

Net effect: every LM Studio embedding came back as 1024 zeros regardless
of the model's true dimension. Qdrant then rejected the upserts with
`Vector dimension error: expected dim: <model>, got 1024`, and indexing
silently failed with all points skipped.

Fix: pass `encoding_format: "float"` explicitly. The SDK detects the
user-provided value (hasUserProvidedEncodingFormat=true), skips the
decode step, and returns LM Studio's float array as-is.

Verified with Qwen3-Embedding-8B (4096-dim): all DEBUG_LMSTUDIO_EMBED
log entries now show firstEmbeddingDim=4096, no skipped upserts.

src/services/provider-lmstudio.ts

@@ -179,9 +179,20 @@ export class LMStudioEmbeddingProvider implements EmbeddingProvider {
     // No `dimensions` parameter: LM Studio doesn't implement Matryoshka projection.
     // The model returns its native dimension and we trust the user to have set
     // EMBEDDING_DIMENSIONS to match.
+    //
+    // `encoding_format: "float"` is REQUIRED. The OpenAI SDK (6.x+) defaults to
+    // `encoding_format: "base64"` for performance, then unconditionally decodes the
+    // response with toFloat32Array(). LM Studio ignores `encoding_format` and always
+    // returns a plain JSON array of floats. The SDK's decode path then runs
+    // `Buffer.from(<array>, 'base64')` — Node.js silently drops the encoding for
+    // array inputs and clamps each float (<1.0) to uint8 0, producing a 4096-byte
+    // zero buffer that gets reinterpreted as a 1024-element Float32Array of zeros.
+    // Setting `encoding_format: "float"` makes the SDK skip the decode step entirely
+    // (see openai-node/src/resources/embeddings.ts: `if (hasUserProvidedEncodingFormat)`).
     const response = await client.embeddings.create({
       model,
       input: texts,
+      encoding_format: "float",
     });
     const sorted = response.data.sort((a, b) => a.index - b.index);
     return sorted.map((d) => d.embedding);