feat(examples): scorecard, held-out-gate, user-simulation-driver

[drewstone]

Summary

Three production patterns the existing six examples did not cover. All three run offline; each is a single index.ts + a short README.md in its own dir.

• examples/scorecard/ — AgentProfile + agentProfileHash + the append-only (persona × profile) JSONL log + diffScorecard (Cohen's d + Welch's t-test verdicts). Demonstrates the one-liner CI guard pattern: diff.cells.filter(c => c.verdict === 'regressed'). Output prints 1 regressed · 1 improved · 2 flat · 0 new and points the regression at the exact cell.
• examples/held-out-gate/ — HeldOutGate.evaluate across all three decision paths: a clean promote, a few_runs rejection, and the classic overfit pattern (search 0.95, holdout 0.55) where the gate correctly refuses to ship. Shows paired CI, rejectionCode, overfitGap.
• examples/user-simulation-driver/ — decideNextUserTurn driven by a scripted TCloud mock so it runs without an LLM. Multi-turn loop with the DONE sign-off. Swap the mock for @tangle-network/tcloud and the same loop drives a real LLM persona.

Test plan

• pnpm typecheck — 0 errors
• pnpm test — 1306 passed (135 files)
• pnpm exec biome check examples/{scorecard,held-out-gate,user-simulation-driver} — clean
• All three runnable: pnpm tsx examples/<dir>/index.ts

[tangletools]

Verified. Three production-pattern examples filling real gaps: scorecard (AgentProfile + diffScorecard, the (#78) build I just shipped — no example existed), held-out-gate (the promotion gate, only previously referenced inside production-loop), user-simulation-driver (decideNextUserTurn with a scripted TCloud mock for offline). All three run end-to-end; typecheck 0, suite 1306 green, biome clean.