feat: add hermetic LongMemEval harness foundation

[wbugitlab1]

Summary

• Add a hermetic benchmark/longmemeval/ foundation for issue feat(bench): LongMemEval-S harness with statistical rigor + CI gate #313 with fixture data validation, six system definitions, manifest hashing/redaction, statistical utilities, markdown table rendering, and a local check entrypoint.
• Add deterministic LongMemEval harness tests and bench:longmemeval:check.
• Document that provider-backed reader/judge runs, real dataset download/submodule policy, historical QA baselines, and real CI benchmark gates remain approval-required future work.

Refs #313

Verification

• corepack pnpm exec vitest run test/longmemeval-harness.test.ts test/eval-adapters.test.ts test/quality-gates.test.ts passed: 3 files / 43 tests.
• corepack pnpm run bench:longmemeval:check passed: { ok: true, fixtureRows: 3, systems: 6, smokeIds: 50 }.
• corepack pnpm run lint passed.
• corepack pnpm test passed after base merge: 212 files / 2920 tests. One earlier full-suite post-merge run hit a transient test/codex-sdk-provider.test.ts 2000ms timeout; the isolated file then passed and the full rerun passed.
• semgrep scan --config p/default --error --metrics=off . passed: 0 findings.
• gitleaks protect --staged --redact passed: no leaks. It initially caught two synthetic redaction-test literals, which were changed to runtime-composed fakes before the passing run.
• OSV was not run because this change does not alter dependencies, lockfiles, container images, vendored code, or third-party package surfaces.