refactor: adopt property_spec() for all PROPERTY_MAPPING specs

[TKaltofen]

Closes #84.

All 259 PROPERTY_MAPPING specs across 66 files move from hand-written dicts to mloda 0.10.0's property_spec() builder. Spec invariants are now validated at import, and every option carries a real explanation, which is what declared_option_keys() and the docs surface to users.

The trap this had to avoid

property_spec() defaults context=True and always writes the key. mloda's parser reads a missing context as False (FeatureChainParser._is_context_parameter). 87 specs (every connector spec) omitted context and were therefore group options. Migrating them naively would have silently flipped them to context, changing feature-group resolution and hashing.

Those 87 sites get an explicit context=False. The context=True count is 172 before and 172 after.

How behavior preservation was verified

Every FeatureGroup's effective spec was snapshotted before and after (allowed_values, default, effective context, effective strict_validation, validators), normalizing each absent key to the parser's own fallback.

Result across 67 classes and 259 specs: zero semantic differences. The only delta is the added explanation strings.

The 54 authored explanations

54 specs carried no explanation (their text lived inside the value dict, e.g. {"sentence": "Sentence-boundary aware chunks"}). These collapse to 13 distinct option keys, because a base and its concretes share an option and differ only in the value space they allow. Each key gets one explanation, used identically by the base and its concretes. The per-value descriptions still live in allowed_values, so nothing was lost.

Also

• New guard test_group_context_split_is_stable pins connector options as group and every other family as context. Nothing locked this before, and it is the single highest-risk failure mode of this refactor. Mutation-checked: it fails if an explicit context=False is dropped.
• demo.ipynb taught the old hand-written dict form under "swap in your own embedder"; it now teaches property_spec.

Verification

tox green: 706 passed, 7 skipped, ruff format, ruff check, mypy --strict (241 files), bandit.

Reviewed independently by two deep reviews (Claude Opus and codex); both confirmed no behavior change. Findings on import ordering and on two leaf explanations that disagreed with their base are fixed in 04d3d64.

Note for reviewers

The registry guide feature-group-patterns/11-options.md is stale and was not used as the source of truth here. It still documents the retired flattened form as valid, and it claims property_spec accepts validation_function and type_validator (both raise TypeError; the real parameters are element_validator and match_guard). This PR was written against the verified 0.10.0 signature. Worth a separate upstream fix.