[codex] add numeric release gates

[WaffleBits]

What changed

• Extended release observations with optional model, backend, accelerator, and numeric output evidence.
• Added model/backend/accelerator-scoped numeric tolerances for cases where backend fingerprints differ but numeric drift is within policy.
• Added schema v2 release reports with numeric drift metrics and per-segment summaries.
• Added a checked numeric-tolerance fixture and CI coverage.

Why

Current inference runtime roles emphasize correctness gates across heterogeneous accelerators, shadow/canary validation, and quantitative evidence for backend regressions. This keeps the runtime repo focused on executable validation evidence instead of profile wording.

Validation

• cargo fmt --all --check
• cargo test --all-targets in WSL
• cargo clippy --all-targets --all-features -- -D warnings in WSL
• cargo run --release replay and all gate fixtures in WSL
• git diff --check