Fix CoreMark benchmark comparison validity

[arthurianresolve]

Summary

• Build the PR and main implementations with the same CoreMark harness and workload.
• Fail when either side cannot produce the pinned fixed-clock workload instead of silently skipping instruction counts.
• Remove the unsupported widest-range-as-noise classification and the unexplained 1% warning.
• Describe Callgrind counts and timed-score ranges without overstating what they prove.

Why

This is a focused follow-up to nasa#124. Its baseline currently predates COREMARK_FIXED_CLOCK, so the PR cannot exercise its own two-sided Callgrind comparison. It also treats the widest range from three timed samples as a regression threshold, although that range is not a confidence bound.

Saving the PR harness before checking out main holds the measurement apparatus constant while varying the implementation under test. Both sides must then produce the known 10.000 fixed-clock score before instruction counts are accepted.

Validation

• Actionlint 1.7.12
• Workflow YAML parsing
• Embedded Bash and JavaScript syntax checks
• cargo fmt --all -- --check
• cargo check -p spacewasm_std --benches
• cargo clippy -p spacewasm_std --benches -- -Dwarnings
• git diff --check
• PR build and current main rebuilt with the saved harness both produced CoreMark Score: 10.000 with four clock_ms calls

Scope

Two workflow files only; no Rust production code changes.

AI usage disclosure

OpenAI Codex was used to analyze PR nasa#124, implement the changes in .github/workflows/ci.yml and .github/workflows/comment.yml, and run the validation listed above. The resulting diff is submitted for author and maintainer review.