fix(search): batch missing-embedding calls to avoid API batch-size limits

[octo-patch]

Fixes #1482

Problem

_extract_embeddings in search_handler.py collects all documents that lack a cached embedding and calls embedder.embed(all_missing) in a single shot. Providers such as Dashscope text-embedding-v4 reject or silently return None when the batch is too large (e.g. 25 documents), which then causes a TypeError when the code tries to iterate over the None result, or silently drops all missing embeddings.

The warning message before this:

[SearchHandler] MMR embedding metadata missing; will compute missing embeddings: missing_total=25

followed by a crash or silent failure in the MMR deduplication path.

Solution

Split missing_documents into chunks of _EMBED_BATCH_SIZE (16) and call embed() for each chunk, extending a combined result list. Batches that return None/empty are skipped gracefully so remaining embeddings can still be used.

_EMBED_BATCH_SIZE = 16

computed: list[list[float]] = []
for i in range(0, len(missing_documents), _EMBED_BATCH_SIZE):
    batch = missing_documents[i : i + _EMBED_BATCH_SIZE]
    batch_result = self.searcher.embedder.embed(batch)
    if batch_result:
        computed.extend(batch_result)

Testing

• Verified with 25 missing documents: previously crashed with TypeError; now completes successfully using 2 batches of 16 and 9.
• Verified with <16 missing documents: behaviour unchanged (single call).

[Memtensor-AI]

Automated Test Results: PASSED

Cloud test-engine rerun against dev-v2.0.22 completed successfully after migrating this run from the old 10010 queue to the updated cloud test-engine.

• Run: tr-99634e82-d49 on cloud test-engine 10011
• Scope: memos_python_core/changed-python-source
• Result: 2 passed, 0 failed, 0 skipped

Manual code review is still required before merge.

[Memtensor-AI]

Thanks for the contribution. This PR addresses the same issue as #2041 (Fixes #1482).

We are going to continue the review in #2041 because it uses a smaller batch size, preserves per-batch index mapping, and includes a regression test for the missing-embedding batching behavior.

To keep the review queue focused, I’m closing this PR as a duplicate of #2041. We can carry any useful hardening ideas from this PR into the #2041 review if needed.