refactor(llm): simplify utils/llm, bounded refine loop, trim tests

[spalen0]

Cleanup + quality pass over utils/llm, plus a test trim.

Simplify / reuse

• Shared error wrapping — the copy-pasted except LLMError: raise / except Exception: raise LLMError(...) ladder (4×) is consolidated into a wrap_llm_errors contextmanager in base.py. Both providers use it; exception types and messages are preserved (incl. the distinct "not valid JSON" message).
• Regex/import hoisting — import re and the eth_utils selector import moved to module scope; the trailing-risk-tag regex is precompiled and the marker pattern is lru_cached (no per-call imports/recompiles).
• Fixed a stale _collect_state_reads docstring.

Quality

• Bounded self-critique loop — _refine_summary now loops up to MAX_REFINE_ROUNDS (=3), stopping early on PASS. Cap is about quality (critique converges in 1–2 rounds; over-editing degrades), not cost.
• refine default-on for explain_transaction / explain_batch_transaction, so every protocol gets the critique pass. The loop runs on the authoritative summary; the detail is still expanded once from the frozen summary, preserving the summary↔detail consistency guarantee.

Tests

• Dropped the brittle LLM/simulation orchestration tests (call-count/plumbing that mock the external API away) and kept the deterministic prompt-building + reply-parsing tests. Net −376 lines.
• Trade-off: the refine-loop and two-stage-consistency behaviors are no longer covered (they're provider-mocked) — deliberate, given we lean on the external API.

Verification

• ruff check + ruff format --check clean.
• pytest tests/test_ai_explainer.py tests/test_llm.py → 85 passed.

Not included (follow-up)

• Best-of-N + judge for a further output-quality bump — discussed, not built. Can add on request.

🤖 Generated with Claude Code