fix(memory): match a Latin word embedded in unspaced-script text

[LHMQ878]

Link to Issue or Description of Change

Related: #5501 (same code path, the Latin-word half of the problem was left open)

Problem:

InMemoryMemoryService.search_memory tokenizes an event with re.findall(r'\w+', text). Japanese and Chinese are written without spaces, so a whole sentence comes back as a single token: 私はPythonでADKを使っています tokenizes to {'私はpythonでadkを使っています'}.

The Unicode fix in b8ea1e8 handles a non-Latin query by also matching it as a substring of the event text, but a Latin query word is still only matched against whole tokens. So a Latin word embedded in unspaced-script text is unreachable — which is exactly how Japanese/Chinese users write about code, product names and identifiers:

TEXT = "私はPythonでADKを使っています"   # "I use ADK with Python"

# on main
query='Python'   hits=0
query='ADK'      hits=0
query='使って'   hits=1   # non-Latin query works (substring fallback)

Solution:

Index an event token that mixes scripts by its ASCII and non-ASCII runs as well as by the token itself (new _extract_searchable_words), so 私はpythonでadkを使っています also yields 私は, python, で, adk, を使っています.

This makes the embedded Latin word a token of its own and is matched exactly, so the partial-word guard the existing tests assert ('thon' must not match Python) still holds — a plain "substring-match Latin words too" fallback would have broken it. Pure-ASCII text produces the same token set as before, so ranking and result bounding are unchanged. Only the event side is affected; query tokenization is untouched, so the match counts used for ranking stay one-per-query-word.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Added six cases to the existing test_search_memory_non_latin parametrization — three that fail on main and three guards:

event text	query	expected
私はPythonでADKを使っています	Python	1 (fails on main)
私はPythonでADKを使っています	adk	1 (fails on main)
我用Python写代码	python	1 (fails on main)
私はPythonでADKを使っています	使って	1 (non-Latin substring still works)
私はPythonでADKを使っています	Java	0 (no false positive)
私はPythonでADKを使っています	thon	0 (partial Latin word still must not match)

On main, the three new positives fail:

FAILED tests/unittests/memory/test_in_memory_memory_service.py::test_search_memory_non_latin[私はPythonでADKを使っています-Python-1]
FAILED tests/unittests/memory/test_in_memory_memory_service.py::test_search_memory_non_latin[私はPythonでADKを使っています-adk-1]
FAILED tests/unittests/memory/test_in_memory_memory_service.py::test_search_memory_non_latin[我用Python写代码-python-1]
3 failed, 12 passed

With the change:

$ pytest tests/unittests/memory/ -q -p no:randomly
92 passed, 2 warnings in 36.15s

$ pytest tests/unittests/tools/test_preload_memory_tool.py -q -p no:randomly
4 passed

$ pytest tests/unittests/tools/test_agent_tool.py \
         tests/unittests/flows/llm_flows/test_agent_transfer_system_instructions.py \
         tests/unittests/cli/test_service_registry.py \
         tests/unittests/cli/utils/test_service_factory.py -q -p no:randomly
106 passed, 1 skipped

pre-commit run --files <changed files> passes (isort, pyink, ruff, addlicense, codespell, ADK compliance checks).

Manual End-to-End (E2E) Tests:

Same store, searched before and after the change:

svc = InMemoryMemoryService()
await svc.add_session_to_memory(session_with_one_event("私はPythonでADKを使っています"))
for q in ["Python", "ADK", "使って", "thon", "Java"]:
    print(q, len((await svc.search_memory(app_name="app", user_id="u", query=q)).memories))

query	before	after
Python	0	1
ADK	0	1
使って	1	1
thon	0	0
Java	0	0

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.