docs(research): COIN benchmark + skwaq gym study and LOCAL COIN Gym design sketch

[rysweet]

Verdict

READY TO MERGE. All six merge-ready criteria are satisfied (evidence below):
criteria 2, 3, 5, 6 directly; criterion 1 (qa-team) is justifiably N/A for a
no-code / no-runtime docs change (executable verification is the CI docs gate);
criterion 4 is green in CI — every required check passed and
mergeStateStatus = CLEAN. Focused, fact-checked, strict-build-clean docs
addition with no behavioral or security surface. No G1/G2/G3 guideline flags apply.

CI — 100% green (0 failures)

All required checks passed on this PR (head eb29acb9); mergeable = MERGEABLE,
mergeStateStatus = CLEAN:

Check	State	Green run
docs (mkdocs build --strict)	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817928571
pre-commit (fmt + clippy --all-targets --all-features --locked + tests)	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817886323/job/85461890718
build	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817928571/job/85462021888
install-real	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817886323/job/85464213548
e2e-dashboard	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817886323/job/85464213538
cargo-audit	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817886323/job/85461891006
cargo-deny	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817886323/job/85461890725
cargo-vet	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817886323/job/85461890804
npm-audit	✅ pass	https://github.com/rysweet/Simard/actions/runs/28817886323/job/85461890858
GitGuardian Security Checks	✅ pass	https://dashboard.gitguardian.com

Merge-ready evidence

1. qa-team / agentic scenarios — N/A (justified). gadugi-test targets
executable app/runtime scenarios; this PR adds no code and no user-facing runtime
behavior. Executable verification is the CI docs gate (criterion 4), which passed.

2. Docs updated (exact changed surfaces). This PR is the documentation:
docs/research/coin-benchmark-and-skwaq-study.md (new) and mkdocs.yml (nav wiring
so the page is discoverable / non-orphaned). No other surfaces touched.

3. Quality-audit — 3 passes, ended clean (zero critical/high; zero medium
correctness/security): (a) fact-check — every leaderboard number, signal-algebra
set, and run command cross-checked against the COIN site/app.js and skwaq
spec/source (hallucinated web-search "COIN" hits discarded); (b) link/structure —
mkdocs build --strict (mkdocs 1.6.1 + mkdocs-material, matching docs.yml,
validation.*: warn → strict errors) built clean in ~10s, zero warnings,
confirmed green by the CI docs run above; (c) scope/self-review — 2-file
docs-only diff, no Rust/behavior touched. Final cycle clean.

4. CI — 100% green. See the table above (links to each green run/job).

5. PR description evidence. This body provides concrete evidence for criteria
1–4 and 6, including the CI run/job links above and the diff stat below.

6. Focused diff. 2 files, +546/-0, docs-only, no unrelated edits
(docs/research/coin-benchmark-and-skwaq-study.md, mkdocs.yml).

Summary

Phase-1 (LEARN COIN) + Phase-2 (STUDY skwaq) research spike for the goal
build-a-local-coin-benchmark-harness-and-a-self-…. Docs-only; no Azure VM and
no external COIN result posting (local/notes only). New page
docs/research/coin-benchmark-and-skwaq-study.md:

• Part 1 — COIN (COde → INput): reachability task (input drives execution to a
  target line, graded by harness replay on an instrumented build — objective,
  contamination-resistant, no judge); OSS-Fuzz + CodeQL construction with the
  G/F/L/C/B signal algebra (frontier (B∪C)\(G∪F∪L) vs G\(F∪L)); reach-rate +
  precision scoring; the published 8-model leaderboard; local run mechanics
  (uv sync --extra llm, coin evaluate, coin show; Python ≥3.12 + Docker + LiteLLM).
• Part 2 — skwaq loop: gym harness + modes, TP/FP/FN scoring, the five-phase
  failure-analyst → overfitting-reviewer (~66% reject) → patch → verify-with-rollback
  cycle, train/holdout gating, schema-backed multi-agent debate, profiles.
• Part 3 — LOCAL "COIN Gym" design sketch: run COIN locally, score vs the
  leaderboard, single-model baseline vs multi-agent team, mirroring skwaq's
  failure-analysis + overfitting-reviewer gating with COIN's self-refreshing fresh
  targets as the anti-overfit oracle. Architecture, components, data model, CLI sketch.

Sources verified directly: COIN https://coin-bench.github.io/ (page + assets/app.js),
github.com/COIN-Bench/coin, huggingface.co/datasets/COIN-Bench/coin; skwaq clone
(Specifications/skwaq-gym-design.md, docs/gym-self-improvement.md,
agents/{failure-analyst,overfitting-reviewer}.md, crates/gym/src/{scoring,improve}.rs).

Tracking / out of scope

Phases 3–5 (azlin VM provision, harness build, iterative self-improve) → #2713.
No Azure VM provisioning, no external result posting, no harness code in this spike.