fix: contain a raising match hook instead of failing every feature

[TKaltofen]

Closes #845.

Both halves of the issue still reproduced on main.

Part 2: one broken matcher poisoned every feature

A FeatureGroup whose match_feature_group_criteria raised took the whole filter loop down, so the exception became the resolution error of whatever feature was requested. Reproduced on main: with a throwaway group whose matcher raises RuntimeError("boom"), resolve_feature("some_unrelated_name") returned error='boom'.

The shared filter seam now contains such a raise as a non-match for that candidate alone, and records it as a match hook near-miss so the reason still reaches the user:

No feature groups found for feature name: 'x'.
Feature group(s) eliminated while matching 'x':
  - BrokenFG (match hook): raised RuntimeError: boom

Containing every raise would have been wrong on its own: it also swallowed framework-owned raises that are deliberately loud, turning a conflicting-reader misconfiguration into a silently wrong resolution. Those are marked with escalate_match_abort and cross the seam untouched, keeping their original type, message and traceback:

• the build_effective_options escalation,
• both already set with different values data-access conflicts,
• the mixed-batch and conflicting pinned-file errors,
• the forwarded-name-mismatch guard.

Raises that are a per-candidate judgment stay contained, and each such site now says so in place, so the rule does not have to be rediscovered. The build-phase fail-closed contract is untouched: a raising compute_framework_definition() still aborts the environment build.

A contained raise also rolls back the writes its matcher made to the shared Feature.options, per candidate, so a half-mutated option set cannot reach the group that wins instead.

Part 1: the test-isolation footgun

The autouse cleanup that reclaims a test's throwaway FeatureGroup subclasses had been copy-pasted into eleven modules, and every new module defining one had to rediscover it. One autouse fixture in tests/conftest.py, backed by tests/registry_isolation.py, now gives every module isolation by default. It collects only when a new subclass actually appeared, and pays a second collection only when the first left something behind, so suite wall time is unchanged.

A structural test asserts the fixture is defined exactly once. That test earned its keep immediately: main landed a twelfth copy while this branch was open, and it caught it.

The two direct-matcher sweeps in test_resolve_feature.py now skip a candidate whose matcher raises, closing the mechanism behind the flake in #868.

Verification

Merged with main after it landed the resolution-module split (#891), the reader option surface (#893) and the gc.freeze() conftest contract (#894). The matcher_error stage and its label moved to whichever module now owns them; the marked raise sites were re-applied against the reshaped base_input_data.py.

tox green on the merged branch: 7567 passed, 170 skipped, ruff format, ruff check, licenses, mypy --strict over 910 files, bandit. CI green on all 17 jobs across 3.10 to 3.14.