[codex] Remove perf GitHub summary

[cquil11]

Removes the perf results table from GitHub step summaries.

Keeps eval summaries intact.

Validation: focused tests + workflow YAML parse.

[claude]

LGTM — clean removal of the perf step-summary path; the inlined constants and load_json in collect_eval_results.py match the deleted summarize.py verbatim, and no other callers remain.

Extended reasoning...

Overview

This PR removes the perf results markdown table from GitHub step summaries. It deletes utils/summarize.py entirely, drops the 'Print summary' step from .github/workflows/collect-results.yml, and inlines the handful of column-name constants plus the load_json helper (previously imported from summarize) directly into utils/collect_eval_results.py — its only importer. Documentation (README.md, AGENTS.md) and a couple of comment/test-name references are updated to match.

Security risks

None. This is a pure removal of a debug/reporting surface — no changes to auth, secrets, execution paths, or data handling.

Level of scrutiny

Low. This is a mechanical refactor: delete a module, inline the ~25 lines of constants and one helper into the sole consumer, and remove a step from a workflow that only wrote to $GITHUB_STEP_SUMMARY. I verified no remaining references to the deleted module (grep for utils/summarize / from summarize / import summarize turned up nothing; the other 'summarize' hit is experimental/CollectiveX/summarize.py, unrelated).

Other factors

The eval summary path (utils/collect_eval_results.py) is preserved, which matches the PR description. The test file rename (SUMMARIZE_KEYS → AGGREGATE_KEYS, test_processor_emits_required_summarize_keys → test_processor_emits_required_aggregate_keys) is cosmetic and keeps the same assertions. Bug hunter found nothing.