[codex] Add inference telemetry release gates

[WaffleBits]

Summary

• add mirrored inference telemetry for model version, queue depth, KV memory pressure, TTFT, decode-token latency, and token-trace fingerprints
• extend release reports with aggregate and segmented telemetry fields
• hold candidates for TTFT, decode-token p95, or memory-pressure regressions while preserving rollback for correctness, numeric drift, and error-rate failures
• add replay pressure summaries for max queued requests, max active requests, peak KV pressure, queued-pressure ticks, active-capacity ticks, and pressure ratios
• add a checked workload-pressure fixture/artifact that completes eight mixed-priority requests in 27 ticks and peaks at 86.666667% KV pressure
• update docs, tests, fixtures, and checked artifacts

Validation

• cargo fmt --all --check
• cargo test --all-targets
• parsed all JSON fixtures/artifacts with PowerShell ConvertFrom-Json
• git diff --check