refactor: retire transitional PROPERTY_MAPPING parser seams (#798)

[TKaltofen]

Retires the transitional PROPERTY_MAPPING parser seams left by the hardening epic (board issue os-005, migrated from #798, phase 7 of the epic tracked in #761).

What changed

1. Trivial private wrappers removed. FeatureChainParser._is_context_parameter, _is_strict_validation, _get_element_validator, and the _extract_property_values alias are gone; production code and tests read the canonical PropertySpec fields and call the public extract_property_values(). The mixin's test-only _build_effective_options pass-through is gone too (no public compatibility contract required it).

2. Presence semantics centralized. The allow_explicit_none presence decision existed as six inline copies (four in the parser, one in the mixin, one in feature_group.py). They are replaced by one shared helper, option_key_is_present, keeping the behavior matrix from feat: opt-in way to distinguish explicit None from an absent option #768: an opted-in explicit None counts as present, a flagless present-as-None does not. Equivalence at every site follows from Options.get/__contains__ semantics and is pinned by tests.

3. Framework-owned exception containment narrowed. check_required_when no longer swallows a raise out of framework-owned build_effective_options as a non-match. No user callback runs inside the build, so such a raise is a framework defect (or a user configuration error carrying actionable guidance) and now surfaces. Containment of user-supplied validators, guards, and required_when predicates is unchanged.

4. Diagnostic replay removed from the engine. The resolution-failure report used to re-run _strict_validation_rejection_reason per candidate, re-executing user validators and parser logic a second time. The engine now opens a recording window around each candidate's real match call and captures the structured rejection reason as it happens (strict value rejections, strict match_guard rejections, name-path missing-required-option reasons), keyed by the candidate class object so same-named classes from different modules keep their own reasons. _strict_validation_rejection_reason remains as a standalone diagnostic facade producing identical messages, but the engine never calls it: user callbacks run exactly once per candidate per resolution.

No public API is removed; all deleted names were underscore-private transitional seams. attribution/ATTRIBUTION.md picked up the current dependency resolution (defusedxml, nltk 3.10.0) per the tox regeneration rule.

Review

Two independent reviews (Claude Opus and codex) ran on the diff. Both flagged the same defect in the first recorder implementation: name-keyed, first-wins recording collapsed same-named candidate classes. Fixed by the per-candidate recording window with class-object attribution, with a regression test. Remaining accepted nits: dropped an unused parameter and documented the deliberately kept string-path containment asymmetry (a config error raised while merging name bindings stays contained on the string path, while the required_when path surfaces it).

Testing

Full tox green twice (after the main change and after the review fixes): 7160 passed, 170 skipped, ruff format/check, license allowlist, mypy --strict (883 files), bandit.