Add Xeon 6973P-C performance profile

Author: satyakigh

src/performance-harness/README.md

@@ -2,7 +2,10 @@
 
 Performance is checked against a versioned environment profile, never against another Git revision.
 
-* `expected/github-ubuntu-x64-amd-epyc-7763.json`, `expected/github-ubuntu-x64-amd-epyc-9v74.json`, and `expected/github-ubuntu-x64-intel-xeon-platinum-8573c.json` are separate tight contracts for CPU models used by GitHub-hosted `ubuntu-latest` x64 runners. The harness derives and enforces the matching model automatically.
+* `expected/github-ubuntu-x64-amd-epyc-7763.json`, `expected/github-ubuntu-x64-amd-epyc-9v74.json`,
+  `expected/github-ubuntu-x64-intel-xeon-platinum-8573c.json`, and
+  `expected/github-ubuntu-x64-intel-xeon-6973pc.json` are separate tight contracts for CPU models used by
+  GitHub-hosted `ubuntu-latest` x64 runners. The harness derives and enforces the matching model automatically.
 * `expected/local-macos-arm64.json` is the contract for the recorded reference Apple Silicon Mac. The harness rejects a different Mac model instead of comparing unlike hardware.
 
 The `check` command spawns the current release executable for both engines across synthetic, real-template, and security workloads. Each case discards its first process launch, then uses the median of five independent launches. An apparent failure receives four additional samples and is evaluated again over the combined set.
@@ -38,6 +41,6 @@ cargo run --locked --release -p performance-harness -- update
 
 GitHub expectations must be measured on GitHub-hosted runners. Every workflow run uploads `performance-candidate-baseline.json`; after a confirmed improvement or an intentional regression, use the candidate to update the matching CPU-specific profile in review. Never copy local Linux measurements into a GitHub profile.
 
-A previously unseen GitHub CPU still fails the required check. Instead of failing before measurement, the harness collects nine independent launches per case and uploads a CPU-enforced candidate plus raw results. Validate repeated hosted-runner evidence before adding the candidate under its suggested deterministic profile name; the unknown hardware cannot pass until that profile is checked in.
+A previously unseen GitHub CPU still fails the required check. Instead of failing before measurement, the harness collects nine independent launches per case and uploads a CPU-enforced candidate plus raw results. Validate it with at least two independent hosted runs of the same performance-relevant revision; baseline-only changes do not invalidate confirmation. The second candidate must pass the first candidate's two-sided contract with no per-case or aggregate failures. Record the workflow, job, revision, and held-out ratios in the profile provenance before adding the candidate under its suggested deterministic profile name. The unknown hardware cannot pass until that profile is checked in.
 
 Changing workloads causes an exact case-set mismatch and requires an intentional baseline regeneration. Environment mismatches fail before measurement rather than producing misleading performance results.