fix: prevent crashes in jieba text_exists and rate_limit flush_cache

[ifer47]

Summary

• Jieba keyword table crash: text_exists() crashes with TypeError when _get_dataset_keyword_table() returns an empty dict {}, because set.union(*{}.values()) requires at least one argument. The method only guarded against None but _get_dataset_keyword_table returns {} at line 184. Fixed by changing the guard from keyword_table is None to not keyword_table.
• Rate limiting TOCTOU race: flush_cache() had a race condition between redis_client.exists() and redis_client.get(). If the Redis key expires between the two calls, get() returns None and calling .decode("utf-8") on it raises AttributeError. Fixed by replacing the exists() + get() pattern with a single get() call and a None check.

Reproduction

Bug 1 — set.union on empty dict:

>>> set.union(*{}.values())
TypeError: unbound method set.union() needs an argument

This is reachable when _get_dataset_keyword_table() returns {}, which happens at line 184 when no keyword table data exists.

Bug 2 — Redis key expires between exists() and get():

# Key exists during exists() check, but expires before get()
cached = redis_client.get(key)  # Returns None
int(cached.decode("utf-8"))     # AttributeError: 'NoneType' object has no attribute 'decode'

Test plan

• Added test for text_exists() with empty dict {} — verifies it returns False instead of crashing
• Added test for flush_cache() when Redis key disappears — verifies it falls back to local value instead of crashing
• Updated existing test_should_sync_max_requests_from_redis_on_subsequent_flush to remove stale exists.return_value mock (no longer needed after fix)

🤖 Generated with Claude Code Best