Union search hangs indefinitely when prefix query token contains 2+ multibyte UTF-8 characters

Description

A multi_search request with union: true hangs forever (never returns a response) when the query token:

1. contains two or more multibyte (non-ASCII) UTF-8 characters, and
2. is a prefix of an indexed token (prefix search enabled, which is the default), and
3. union: true is set.

The exact same query returns instantly as a regular (non-union) multi_search, and also returns instantly with prefix=false. Tokens with only one multibyte character work fine. The server stays healthy while the request hangs — only the affected request spins (looks like an infinite loop in the union prefix-expansion path).

We hit this in production with Danish food searches, e.g. a user typing blåbæ on the way to blåbær (blueberry).

Steps to reproduce

Fresh single-node Docker container, no other data:

docker run -d -p 8108:8108 typesense/typesense:31.0.rc10 --data-dir /tmp --api-key=test123

curl -s "http://localhost:8108/collections" -X POST \
  -H "X-TYPESENSE-API-KEY: test123" -H "Content-Type: application/json" \
  -d '{"name": "repro", "fields": [{"name": "name", "type": "string"}]}'

curl -s "http://localhost:8108/collections/repro/documents" -X POST \
  -H "X-TYPESENSE-API-KEY: test123" -H "Content-Type: application/json" \
  -d '{"name": "blåbær"}'

# This request never returns:
curl -s "http://localhost:8108/multi_search?page=1&per_page=10" -X POST \
  -H "X-TYPESENSE-API-KEY: test123" -H "Content-Type: application/json" \
  -d '{"union": true, "searches": [{"collection": "repro", "q": "blåbæ", "query_by": "name"}]}'

Expected behavior

The union search returns the prefix match blåbær, like the non-union search does.

Actual behavior

The request hangs indefinitely (we observed >60s in production until our proxy killed the connection; >15s locally before aborting client-side). /health keeps returning {"ok": true} the whole time.

Characterization matrix

Same collection as above (docs: blåbær, blåbærsyltetøj, jordbær, røræg, chicken), all with union: true unless noted:

Multi-token queries where only the last token is the prefix (e.g. blåbær med flø, 1 multibyte char in last token) also work.

Environment

• Typesense 31.0.rc10 (also reproduced on 31.0.rc3)
• Official Docker image, single node, default settings
• Reproduced both on a clean local container and on our production cluster

Reproduce the hang using the Docker command and curl request in the issue, then trace the union=true multi_search prefix-expansion path for UTF-8 tokens. No source file or test is named; done means the blåbæ prefix returns blåbær promptly under union search, while the listed non-union, prefix=false, and single-character cases remain working.