refactor(tests): split llm_tests — plumbing (actions + pool) into own file

`src/tests/llm_tests.rs` was at 903 lines.  The prompt / instruction /
music-theory / style / dsp-math coverage (LLM *content* tests) and the
action-extraction / json-repair / pool / agent-model-switch coverage
(LLM *plumbing* tests) are independent enough that they don't share
fixtures — natural cut.

Moved the plumbing mods (action_extraction, json_repair_tests,
pool_tests, agent_model_tests) to a new `src/tests/llm_plumbing_tests.rs`
(443 lines).  llm_tests.rs drops to 428 lines; what's left stays
focused on prompt construction, instruction matching, style catalogs,
and the couple of DSP / music-theory sanity checks.

987 lib tests pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Tok