docs(examples): simplify + sharpen the example set + clean up the coding benchmark

[drewstone]

Applies the real DX wins from an example-set + coding-benchmark review. Every change cuts duplication or restatement prose; no behavior change. The load-bearing properties of the coding benchmark are preserved and execution-verified: the held-out test-execution anti-cheat, the firewall (held-out tests never in the box during the turn), the stats power-floor guard, and the catalog/docs gate.

examples set

• README: a 3-example Quickstart (driver-loop / supervise / improve) now leads, above the full 6-tier table — the "run these first" path is no longer buried.
• strategy-suite + strategy-evolution: the verbatim ~60-line counterEnv fixture (token-bucket counter Environment) is extracted to strategy-suite/counter-env.ts and imported by both; each example now shows only its distinct concept (compare strategies vs. the holdout gate).
• fleet-delegation: renamed UPPERCASE module-globals (SIBLING_ENV/FLEET_ENV → siblingEnv/fleetEnv) — models the documented publish-gotcha house rule instead of teaching the banned pattern.
• self-improving-loop + supervisor-loop/shared: stop re-teaching the shot/round vocabulary in 3 places; point at driver-loop (the canonical home). The "do NOT mistake this scripted brain for the pattern" warning is kept.
• supervisor-loop/README: leads with the one-knob sentence + a 3-row table (runner → backend → command); the per-backend exposition is demoted under a ## Details section.
• agents-of-all-shapes/README: the Python agno snippet is now flagged as illustrative — not run by run.ts, not typechecked.

coding benchmark — 9 files, 2067 → 1978 LOC (file count unchanged)

• delete the dead wilcoxon/PairedTest path in stats.ts (no caller or CLI flag ever passed 'wilcoxon'); always run the paired t-test. Dropped the dead import + the README/comment mentions.
• rename RunArtifactSummary → BenchmarkSummary — it counts {records, leaderboard}, unrelated to the per-cell RunArtifact the judges score.
• de-dup the firewall / scoring-order / offline-degeneracy prose to one canonical home each (the dispatch.ts firewall banner, the README tables); the other copies become one-line pointers.
• profiles.ts: drop restatement comments (// NO prompt, NO resources etc.), keep the two non-obvious facts (metadata.harness carries the selector; the snapshot-date requirement).
• note that score()'s field order is searchScore-only in this example.

skipped (with reason)

• intelligence-recommend importing fixtures from improve.ts — improve.ts runs main() as an import side-effect, so importing it would execute the improve pipeline twice when running intelligence-recommend; sharing would require guarding main() behind import.meta.url + exporting the pieces, which is more churn than the ~26 trivial fixture lines it would save, and it costs each example's self-containedness. Not worth it.

verify (all green, offline, no creds)

• pnpm run build ✓ · pnpm run typecheck (src + examples) ✓ · pnpm run lint (328 files) ✓
• pnpm run docs:check (catalog regenerate + git diff --exit-code docs/api + freshness) ✓ — no drift
• coding-benchmark + supervisor-loop example tests: 10/10 pass
• offline pnpm tsx examples/coding-benchmark/benchmark.ts ✓ — power-floor NOTE + SIGNIFICANT-suppression intact
• anti-cheat gap intact (execution-verified): rate-limiter cheat held-out 0.50 → composite 0.590; real held-out 1.00 → composite 0.940; gap +0.350
• reps don't fake n (execution-verified): reps=1 vs reps=3 leave the leaderboard CI width unchanged (0.5000/0.3000), n stays = distinct scenarios

DO NOT MERGE — operator review.

[tangletools]

✅ Auto-approved PR — 1bdbe4ec

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-24T13:21:26Z}