refactor(memory): modularize benchmarks and experiments

[drewstone]

Summary

• split benchmark, experiment, and improvement internals behind unchanged public entry points
• split oversized test files and centralize shared fixtures
• remove the holdout/types import cycle
• prepare package version 4.0.1

Compatibility

• public declarations are unchanged: root 512/512, memory 96/96, benchmarks 52/52
• original implementation declarations are preserved: benchmark 107/107, experiment 58/58, improvement 41/41, holdout 23/23
• original test cases are preserved: memory 110/110, benchmarks 30/30
• no source import cycles remain

Checks

• pnpm lint
• pnpm typecheck
• pnpm test: 424 passed, 10 intentionally skipped
• pnpm build
• pnpm verify:package: clean install, five imports, CLI version, and repack
• deterministic Mem0 deletion test passed 10/10 repeated runs

[tangletools]

✅ Auto-approved drewstone PR — ddc7e3d8

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-18T14:41:09Z}