Normalize tuple defaults on list fields instead of warning

[pufit]

Symptom

Since #205 (ced5afa, the list-field coercion), every config load logs:

[WARNING] nerve.coerce: LangfuseConfig.redact_patterns: ignoring non-list config value ('sk-ant-…', …), leaving it as-is

on any config that doesn't mention langfuse.redact_patterns — i.e. for most installs, on every nerve command and session.

Root cause

_DEFAULT_LANGFUSE_REDACT_PATTERNS is a module-level tuple, and LangfuseConfig.from_dict hands it through as the field default. _str_list deliberately passes non-lists untouched (so coerce_scalars can wrap a ${VAR} string as one element), but _scalar_to_list only knew None/str/scalars — a tuple fell through to the warn-and-skip branch. Harmless in effect (the consumer list()s the field, all four patterns still compiled), but pure noise on every load.

Fix

• config.py: materialize the default as list(...) in from_dict, so the field arrives as its declared type. Same None-vs-[] semantics as before (a bare redact_patterns: keeps the defaults, an explicit [] turns redaction off).
• coerce.py: teach _scalar_to_list to copy tuples element-wise. YAML never parses to a tuple, so one can only be a code-side default — the elements are already the intended entries, only the container is wrong.

Tests

• A defaults sweep: from_dict({}) must yield an actual list for every declared list[X] field across all config dataclasses — the existing sweeps only fed values in and never checked what a builder produces when the section is absent, which is exactly how this slipped through. (DEFAULT_CLAUDE_MODELS is the only other tuple constant near a list field and is already list()-wrapped at its use site; the sweep now pins that class of bug.)
• A targeted regression test: an absent langfuse block keeps the default redact set with zero nerve.coerce warnings.

Both fail on main, pass here. Full suite: 2940 passed.

Generated by Nerve