feat(optimization): unify method admission and verify search evidence

[drewstone]

Complete optimization methods could reach final assessment through selfImprove without the history admission available in compareOptimizationMethods. Both workflows now share detached method inputs, result validation, cost reconciliation, and optional history admission. Defaults remain compatible; strict ledger verification resolves bytes through campaign storage, checks their existing content identity, and replays the canonical journal before final dispatch.

Optional final-measurement evidence binds the actual baseline and selected winner to existing EvidenceReceipt records. Callers supply evaluator, environment, and authority; Eval derives input, output, and measurement identities. This does not rerank the optimizer's winner or infer independent authority. The release versions for npm and Python are aligned at 0.176.0, with generated API census and maintained benchmark pins updated.

Validation:

• Full JavaScript suite: 5,804 passed, 3 skipped (396 files passed, 2 skipped).
• Strict history and measured receipt tests cover both workflows, missing/incomplete/tampered history, invalid chains, replay mismatches, exact output identity, and explicit authority.
• Typechecks for library, examples, and scripts; lint; build; package verification; built public import smoke passed.
• Official GEPA/SkillOpt TypeScript integrations: 2 passed. Python source-optimizer suite: 207 passed, 3 cases skipped, and 1 module skipped (DSPy and published-GEPA checks use separate CI environments).

No model-backed paid calls were made. Merge and publication remain coordinated with the shared-stack release.