docs(examples): tighten the good/great examples — simpler, zero capability lost

[drewstone]

Tightens the GOOD/GREAT examples a prior three-persona review flagged as marginal-but-left-alone. Every change makes the example simpler / clearer / more reusable while keeping exactly what it does and teaches — fixtures are extracted (not deleted), concepts de-duped (not dropped), an offline mode added (not removed). Verified each touched example still typechecks, runs, and keeps its output/teaching.

What was tightened (per example)

1. driver-loop — extracted the ~70-line worker/adapter/validator fixture to scripted-worker.ts so the driver (the subject) leads. The fold demo + the THE FOLD, PART 1/2 anchors are unchanged. Subject file 211 → 146 lines (+81-line sibling). Output identical.
2. researcher-loop — extracted the ~100-line hand-written fixture to synthetic-researcher.ts so the runLoop + inline-fanout wiring leads. The namespace-leak hard-fail demo is intact (the leak candidate is still pruned, the in-namespace candidate wins). Subject file 201 → 78 lines (+147-line sibling). Output identical.
3. recursive-supervisor — extracted the inline executor-registry plumbing + as-casts to inline-executor.ts so the scope.spawn conserved-budget lesson leads. Part 1 (raw scope.spawn), Part 2 (fanout combinator), and the rejected-third-spawn budget-exhausted demo all kept. Subject file 205 → 131 lines (+94-line sibling). Output identical.
4. strategy-suite — ADDED an offline mode: an in-process mock OpenAI-compatible router (mock-router.ts) gated on the absence of TANGLE_API_KEY. The whole comparison now runs end-to-end with zero credentials, driving the REAL runBenchmark machinery (Supervisor, conserved budget pool, every strategy, the env's own check) against localhost. The live Tangle router stays the drop-in upgrade. Capability added, none lost.
5. chat-handler — the hand-rolled NDJSON drain stays (agent-runtime exports no decoder — checked), but a note now marks it illustrative and points at handleChatTurn as the framing owner so consumers don't think they must re-implement it. Lifecycle envelope demo unchanged.
6. intelligence-recommend — now imports the shared scaffolding (DemoScenario / scenarios / agent / judge / scriptedWinner / profile) from improve.ts instead of duplicating ~50 lines, so it reads as "improve, but the findings come from a recorded trace" — its actual new idea (recordTrace → findings → improve) is all that's left. improve.ts's main() is guarded with an entry-check so the import doesn't double-run. 134 → 88 lines. Both outputs identical.
7. stream-backends — the inline SandboxEvent is now the real exported union from @tangle-network/sandbox (a status the SDK adds is a compile error here, not a silent miss). The OpenAI-compatible backend now defaults to the real OpenAI endpoint so the OPENAI_* env names are honest, with a note that you can point it at the Tangle router. All three transports kept.
8. sanitized-telemetry-streaming — collapsed the near-duplicate safe/verbose halves into one drain(label, collector) helper. Both demonstrated states are kept: default-redaction (PII/args/result/uri stripped) and opt-in verbose (those fields exposed). The task.intent PII footgun note stays.
9. runtime-run — dropped the cosmetic task.metadata workspaceId/threadId duplication of the explicit startRuntimeRun args (these never flowed into the row's typed columns — verified in src/runtime-run.ts). Persisted row identical.

Skipped (no no-capability-loss simplification available)

• supervisor-loop / mcp-delegation / fleet-delegation / agents-of-all-shapes — left as-is per the brief. Their review flags were already addressed, or the "fix" would ADD ceremony (e.g. a 40-member fake SandboxInstance helper). No change made deliberately.
• chat-handler decoder swap — agent-runtime exports the NDJSON producer/framing side (handleChatTurn), not a consumer-side decoder, so there is nothing to import; took the lighter "add a note" path the brief allows.

Verification

• pnpm run build, pnpm run typecheck (src + typecheck:examples), pnpm run lint, pnpm run docs:check — all green (docs freshness: no drift; 0 errors).
• Ran every materially-changed example offline; each keeps its exact output/teaching.
• strategy-suite confirmed to run end-to-end with NO creds (clean exit, mock server closes via finally).
• Merges cleanly into origin/main (git merge-tree).

[tangletools]

✅ Auto-approved PR — 7f19b4b4

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 · 2026-06-24T15:35:59Z}