feat(evals): add an offline TypeSafe/Jev routing replay harness #882

I plan to implement this and will link a draft pull request shortly.

Problem

The TypeSafe/Jev work in #723 and #762 includes useful benchmark and stability evidence, but the evidence is not stored in a versioned, repeatable form. Reviewers and operators cannot replay routing decisions, compare candidate-order behavior, or evaluate a threshold policy without making new provider calls and potentially repeating completion-model spend.

Proposed scope

Add a small offline-first evaluation harness that:

• reads versioned, synthetic JSON fixtures;
• records fixture provenance, candidate descriptions, candidate orders, raw probability distributions, expected target outcomes, and optional per-target cost;
• deterministically recomputes aggregate probabilities, selected targets, confidence, and fallback decisions;
• reports selection counts, order sensitivity, fallback counts, quality, cost, and regret against fixed-target baselines;
• runs without credentials, network access, or provider SDK dependencies;
• documents how to create sanitized fixtures without including credentials, private prompts, or provider-controlled error bodies.

The initial pull request will include representative synthetic fixtures, focused tests, and usage documentation.

Out of scope

• Implementing or changing the runtime TypeSafe/Jev router in feat(routing): compare configured TypeSafe candidates #762.
• Calling TypeSafe or completion-model providers from the default evaluator.
• Claiming that the included synthetic fixtures establish production routing quality.
• Committing real customer prompts, credentials, or private provider responses.

Acceptance criteria

• The fixture format is explicitly versioned and rejects unsupported versions.
• Replaying the same fixture produces byte-for-byte stable machine-readable output.
• Candidate-order sensitivity and invalid probability data are covered by tests.
• Quality and cost comparisons use declared fixture outcomes rather than live calls.
• The evaluator is entirely offline by default.
• Repository lint, type checking, tests, and formatting checks pass.

This provides the reproducible evaluation boundary suggested in #723 while remaining independent of the runtime implementation under review in #762.