- Use local N = 'llm_search_context' idiom instead of constant string reuse
- Replace rspamd_logger.debugm with lua_util.debugm (rspamd_logger has no debugm method)
- Use extract_specific_urls instead of task:get_urls()
- Add task parameter to query_search_api for proper logging and HTTP requests
- Remove retry logic using non-existent rspamd_config:add_delayed_callback
- Simplify to single HTTP attempt with graceful failure
- Remove retry_count and retry_delay options from config

- Replace manual Redis operations with lua_cache API for better consistency
- Use messagepack serialization and automatic key hashing
- Fix Leta Mullvad API URL to /search/__data.json endpoint
- Add search_engine parameter support
- Remove redundant 'or DEFAULTS.xxx' patterns (opts already has defaults merged)
- Add proper debug_module propagation throughout call chain
- Improve JSON parsing to handle Leta Mullvad's nested pointer structure

- C code (message.c): collect top CTA URLs per HTML part by button weight,
  store in task mempool variable "html_cta_urls"
- Lua API (lua_task.c): add task:get_cta_urls([max_urls]) method
- llm_search_context: use new API instead of reimplementing CTA logic in Lua
- Benefits: single source of truth for CTA logic, uses C knowledge of HTML
  structure and button weights, cleaner Lua code

This provides proper architecture where C code handles HTML structure analysis
and Lua adds domain filtering (blacklists, infrastructure domains, etc.)

Also refactor all absurdic logic it has previously

Expose rspamd_cryptobox_fast_hash via Lua API to allow hash-based
URL deduplication without string conversion overhead. This is critical
for handling messages with large numbers of URLs (100k+) where tostring()
would fill the Lua string interning table and cause memory issues.

Returns 64-bit hash as Lua number, using the same hash function as
internal URL storage for consistency.

Change cache_key_prefix from 'gpt_search' to 'llm_search' to align
with the module's current name (llm_search_context).

HTML_DISPLAYED URLs are phishing bait text (display-only) and should
not be considered for CTA (call-to-action) detection. Only real link
targets should be analyzed for CTA purposes.

…ased deduplication

Replace tostring() with url:get_hash() throughout URL extraction to avoid
filling the Lua string interning table. Critical for handling malicious
messages with 100k+ URLs where each tostring() would create an interned
string causing memory exhaustion.

Key changes:
- Use dual data structure: array for results + hash set for O(1) dedup
- Add max_urls_to_process=50000 limit with warning for DoS protection
- Track url_index for stable sorting when priorities are equal
- Fix CTA priority preservation: prevent generic phished handling from
  overwriting CTA priorities which include phished/subject bonuses
- Add verbose flag to test suite for debugging

This ensures memory usage is strictly bounded regardless of malicious
input while maintaining correct URL prioritization for spam detection.