docs(coin-gym): reference the reproducible baseline-vs-team measurement

[rysweet]

What & why

The COIN Gym harness (Phases 1/2/4/5) is merged to main, but the goal's
done-criterion — an empirical baseline-vs-team measurement — was not yet
recorded. This PR runs the merged harness end-to-end and records its
reproducible single-model-baseline vs. multi-agent-team result as a
durable reference doc.

Result (bundled sample target set, offline scaffold / mock oracle):

Arm	Reach	Precision	Histogram
baseline (single model)	60.0% (3/5)	60.0% (3/5)	R:3/W:2/A:0/T:0/N:0/E:0
team (multi-agent)	60.0% (3/5)	100.0% (3/3)	R:3/W:0/A:2/T:0/N:0/E:0

The team measurably beats the baseline on precision (+40 pts) at equal reach.
The abstention gate abstains on the two low-confidence targets (liboqs-kem-88
conf 0.44, zstd-huf-1207 conf 0.35) where the baseline submits wrong inputs.
Reproduced end-to-end at commit ad4c8032 (main tip, 2026-07-08).

Durable-documentation (G4) note

The only added doc — docs/research/coin-gym-baseline-vs-team-measurement.md —
is a unit-test-pinned durable reference (numbers asserted by
coin_gym::tests_cli::execute_run_baseline_vs_team_shows_precision_tradeoff),
not a point-in-time report: a future strategy/fixture change is expected to update
it. Its title surface carries none of the report-kind markers checked by
scan_no_point_in_time_report_docs, and it lives under docs/research/ (not a
reserved report dir). The point-in-time milestone finding (numbers + commit SHA)
is offloaded to tracking issue #2823:
#2823 (comment)
(docs/research/coin-benchmark-phase1.md is a pre-existing file on main, not
in this diff.)

Merge readiness

QA-team evidence

• Scenario files: none — docs-only change; no new runtime/user surface for a
  gadugi-test scenario to exercise (internal-only justification below).
• Validation command: mkdocs build --strict (doc build) and
  cargo test -p simard --lib coin_gym::tests_cli::execute_run_baseline_vs_team_shows_precision_tradeoff
  (pins the exact reference numbers the doc records).
• Validation result: passed — mkdocs build --strict exit 0 (no
  warnings/errors); test result: ok. 1 passed.
• Run command: coin-gym run "Claude Opus 4.6" --strategy baseline and
  ... --strategy team on the bundled sample_snapshot.json.
• Run target: local (offline scaffold, mock oracle).
• Run result: passed — baseline reach/precision 60.0%/60.0%
  (R:3/W:2), team 60.0%/100.0% (R:3/W:0/A:2); numbers reproduced in the doc.
• Evidence location: PR result table above + issue COIN Gym Phase 3 — provision azlin VM (DefenderATEVET17) + smoke coin evaluate #2823 comment (linked).
• Rationale for no gadugi scenario: the changed surfaces are three docs +
  mkdocs.yml nav only; there is no code/CLI/config/API behavior change, so the
  authoritative validators are the strict doc build and the existing
  number-pinning unit test, both green.

Documentation

• User-facing docs impact: yes (this PR is the doc).
• Updated docs: added docs/research/coin-gym-baseline-vs-team-measurement.md;
  edited docs/howto/run-the-coin-gym-harness.md,
  docs/research/coin-benchmark-and-skwaq-study.md, and mkdocs.yml (nav).
• PR description links added: new doc registered in the Research nav and
  cross-linked both ways (how-to ⇄ measurement, study ⇄ measurement).
• Rationale if not applicable: n/a.

Quality-audit

• Cycle 1 summary — numerical/claim accuracy. SEEK: could any figure drift
  from source of truth? VALIDATE: re-derived every number from a fresh
  coin-gym run (baseline 60/60, team 60/100; family splits frontier
  50/50→50/100, non-trivial 66.7/66.7→66.7/100; histograms R:3/W:2 and
  R:3/W:0/A:2), the fixture confidences (0.82/0.71/0.66/0.44/0.35 from
  sample_snapshot.json), and the leaderboard figures (30.0/52.5 from
  leaderboard.rs). FIX: none — all matched exactly.
• Cycle 2 summary — link/anchor/build integrity. SEEK: any broken link,
  orphan page, or missing reciprocal link? VALIDATE: resolved all 3 relative
  .md links to existing files, confirmed reciprocal links in how-to/study/nav,
  ran mkdocs build --strict (exit 0, no warnings/errors). FIX: none.
• Cycle 3 summary — G4 durability + no-overclaiming. SEEK: does the doc
  overclaim a live capability result or trip the report-doc scan? VALIDATE: title
  surface has zero report-kind markers; explicit offline scaffold / LOCAL-ONLY
  caveats, an "illustrative only" leaderboard note, and a "what this does and does
  not establish" section scoping the result to offline control-flow. FIX: none.
• Additional cycles: none needed.
• Final clean cycle: cycle 3 — zero critical/high; zero medium
  correctness/security findings.
• Fixes followed default-workflow: n/a (no fixes were required; all cycles clean).
• Convergence summary: three independent audits (accuracy, links/build, policy)
  each ended clean with no findings, so the audit is complete. Landed as commit
  f4a4e935 (docs-only).

CI

• Checks command: gh pr checks 3152.
• Result: all green — build, pre-commit, install-real, e2e-dashboard,
  cargo-audit, cargo-deny, cargo-vet, npm-audit, GitGuardian; 0
  failures, 0 pending. Runs:
  https://github.com/rysweet/Simard/actions/runs/28941803288 ·
  https://github.com/rysweet/Simard/actions/runs/28941803054 ·
  https://github.com/rysweet/Simard/actions/runs/28941773879
• Skipped checks: none.
• Flaky reruns performed: none.
• Real failures fixed: none (green on first run). Local pre-commit
  (rust-only-gate, cargo fmt --check, clippy --release) and pre-push
  (clippy --all-targets --all-features --locked, race test subset) also Passed.

Scope

• Changed files reviewed: gh pr view 3152 --json files → exactly 4 files
  (1 added doc + mkdocs.yml + 2 doc edits). git diff --stat: 3 files changed,
  10 insertions(+), 1 deletion(-), plus the new doc.
• Unrelated changes: none — docs-only; no code changes; every edit is the new
  reference doc or a link/nav entry pointing to it.

Verdict

• Merge-ready: yes
• Remaining blockers: none

Explicit call: this PR is READY TO MERGE (it is not a draft and is not
blocked). Rationale: all six criteria above are satisfied with concrete evidence
— QA validated via mkdocs build --strict + a passing number-pinning unit test
(docs-only, internal-only justification given), the new reference doc is
registered and cross-linked, three independent quality-audit cycles ended clean
(zero critical/high/medium), CI is 100% green with run links, and the diff is
focused with no unrelated edits. Advisory G4 is addressed (durable
unit-test-pinned reference; point-in-time milestone recorded on issue #2823);
live leaderboard grading is separately tracked as Phase 3 (#2823) and does not
gate this docs-only PR.

LOCAL-ONLY: nothing is submitted externally or entered on any leaderboard.