fix(flags): canonicalize numeric group keys for local bucketing

[Kayvan-Zahiri]

Motivation and context

Numeric group identifiers currently reach the local rollout hash through Python string interpolation. For a group flag called group-beta at 50% rollout, group key 0.00001 hashes as 1e-05, returning false, while its canonical string key "0.00001" returns true. Invalid group keys such as booleans or containers can also return a definitive local result instead of leaving remote evaluation eligible.

Normalize group keys before bucketing in both legacy flag-level aggregation and mixed condition-level aggregation. Strings stay unquoted, finite numbers use the existing backend-compatible JSON rendering, and unsupported values raise the evaluator's existing inconclusive signal. No public method, option, or type is added.

This follows the published group-key bucketing requirement.

Verification

The new tests use the public Client.get_feature_flag() and Client.evaluate_flags() APIs with locally loaded definitions. They cover legacy/mixed aggregation, finite numeric/string controls, unsupported boolean/null/container/non-finite values, and remote fallback. Before the fix there are 16 failing subtests; after it all pass.

Focused flag suites: 293 passed, 37 subtests passed. Full SDK suite: 4,620 passed, 16 skipped, 78 subtests passed (134 warnings). Repository-wide Ruff formatting/lint, mypy across 300 source files, and the public API snapshot check pass. The local SDK was exercised against its own evaluator and mocked remote responses; no hosted production flag configuration was changed.

A patch changeset is included. No public API snapshot changes are expected.

Agent context

Autonomy: Fully autonomous

Codex independently selected and reproduced this defect while exercising the Python SDK, implemented the scoped fix, and ran regression and quality checks. A human authorized contribution work but has not reviewed this implementation. The PR is a draft and requires human review; it will not be auto-approved or merged.