chore: mature run_model_sweep.sh into modular, configurable benchmark suite

[AccessiT3ch]

Context

scripts/run_model_sweep.sh was authored as a Study 2a research artifact and works, but it is brittle for future studies:

• Variants are hard-coded as if/elif blocks — adding Study 2b variants requires editing the script
• No test coverage
• No --output-dir override — always writes to the study-id-derived path
• No --timeout override per variant
• The RAM filter list is a static array, not driven by adaptive_k_selector.py

Acceptance Criteria

• Variants loaded from a config file (data/sweep-variants.yml or similar) — adding a new variant requires no script edits
• --timeout and --output-dir CLI flags
• RAM filter driven by OLLAMA_RAM_LIMIT_GB env var or --ram-limit flag
• tests/test_run_model_sweep.sh or equivalent (bats, shunit2, or Python subprocess mock)
• --dry-run output is machine-readable (JSON or structured text)
• Documented in .github/skills/rag-rapid-research/SKILL.md § 5.2

References

• PR feat(rag): Study 2a — local model benchmark infrastructure, sweep, and synthesis #36 Copilot user review: "needs tests, better inline documentation, and a followup issue to mature this into a robust and modular and configurable suite"
• .github/skills/rag-rapid-research/SKILL.md
• scripts/adaptive_k_selector.py

[github-actions]

This issue has been inactive for 60 days and has been marked as stale. If it is still relevant, please comment or remove the stale label.