Enable Thinking Mode before you answer.

You are generating router-v3 training data for Craftly Robot.
Return ONE JSON object only as plain text in your message body (no markdown, no prose, no code fences).
Do NOT create files, attachments, downloads, or tool artifacts.
Do NOT say "saved as .json" or "attached file" - print the full JSON inline in the response.

Objective:
Generate one compact, clean batch for all 3 tracks in one output:
1) intent identifier model
2) conversation model
3) coding model

Behavior architecture to encode in the dataset:
- First principles reasoning (break down problem, verify assumptions, avoid destructive shortcuts).
- Relentless iteration framing (failures treated as signal, attempt-based improvement).
- Anti-hallucination rigor (explicitly request missing context instead of guessing).
- Nuanced emotional intelligence (neutral + uncertainty-aware personal guidance).
- Ethical nudge (explain risk and bridge toward safer alternatives).
- Cost awareness (pause/continue framing with quota or budget awareness).

NodeJS + Flutter best-practice requirements for coding samples:
- NodeJS/TypeScript: strict typing, input validation, async error handling, bounded retries/timeouts, no secret leaks.
- Flutter/Dart: null-safety, immutable state where possible, analyzer-clean code, clear widget boundaries, no anti-pattern side effects.
- Prefer readable production-grade structure over hacks/stubs.

Reference quality examples (for style only, do not copy verbatim):

```typescript
// NodeJS/TS quality example
import { z } from "zod";

const payloadSchema = z.object({ id: z.string().uuid() });

export async function handler(raw: unknown) {
  const payload = payloadSchema.parse(raw);
  return { ok: true, id: payload.id };
}
```
```dart
// Flutter quality example
class ProfileState {
  final String name;
  const ProfileState({required this.name});

  ProfileState copyWith({String? name}) => ProfileState(name: name ?? this.name);
}
```

Return this exact schema:
{
  "version": "craftly-annotation-v1",
  "intent": [
    {
      "user_message": "string",
      "intent": "coding | conversation",
      "route_class": "coding_generate | coding_explain | conversation_general | tooling_or_ops | unsafe_or_policy_block",
      "confidence": 0.0,
      "reason": "short explanation"
    }
  ],
  "conversation": [
    {
      "context": "user prompt",
      "target": "assistant response",
      "style": "concise | balanced | detailed"
    }
  ],
  "coding": [
    {
      "context": "coding request",
      "route_class": "coding_generate | coding_explain",
      "language": "typescript | javascript | python | dart | sql | yaml | json",
      "difficulty": "easy | medium | hard",
      "allowed_roots": [
        "backend/src/",
        "shared/openapi.yaml",
        "shared/generated/",
        "frontend/lib/",
        "frontend/pubspec.yaml",
        "frontend/android/app/src/main/AndroidManifest.xml"
      ],
      "target": "for coding_generate: FILE_BLOCKS_V2 only; for coding_explain: plain explanation only"
    }
  ]
}

Hard constraints:
- Strictly valid JSON. No trailing commas. No comments.
- Response format must be raw inline JSON text only:
  - first character must be "{"
  - last character must be "}"
  - no extra text before or after JSON
- Do not create/write/upload any .json file.
- Every string value must be valid JSON-escaped text:
  - escape all inner double quotes as \" inside string values
  - preserve newlines as \n inside string values
- Never output markdown links, e.g. [FILE:...](...), inside JSON.
- Exactly 1 row in each array: intent, conversation, coding.
- Intent/router consistency:
  - intent=coding => route_class must be coding_generate or coding_explain.
  - intent=conversation => route_class must be conversation_general/tooling_or_ops/unsafe_or_policy_block.
- Coverage requirements for this micro-batch mode:
  - no per-submission route/label coverage requirement beyond schema validity.
  - route/label diversity is tracked across many submissions, not in one single submission.
- Coding coverage:
  - with 1 coding row, route_class may be either coding_generate or coding_explain.
- coding_generate target must be FILE_BLOCKS_V2 only:
  <<<FILE:relative/path>>>
  ...content...
  <<<END_FILE>>>
- FILE_BLOCK markers must be exact:
  - valid start marker: <<<FILE:backend/src/example.ts>>>
  - valid end marker: <<<END_FILE>>>
  - invalid markers (DO NOT USE): <<FILE:...>>, <<[FILE:...](...)>>, markdown links, code fences
- Example of a VALID JSON string in coding.target:
  "target":"<<<FILE:backend/src/config.json>>>\n{\n  \"port\": 3000,\n  \"logLevel\": \"info\"\n}\n<<<END_FILE>>>"
- coding_explain target must NOT contain FILE_BLOCKS_V2 markers.
- Allowed paths for coding_generate are only:
  backend/src/
  shared/openapi.yaml
  shared/generated/
  frontend/lib/
  frontend/pubspec.yaml
  frontend/android/app/src/main/AndroidManifest.xml
- No secrets, credentials, private keys, access tokens, or personal data.
- No harmful/malicious coaching. Refuse unsafe payloads by routing to unsafe_or_policy_block where appropriate.
- Keep data realistic and diverse across backend/frontend/mobile/devops/product contexts.
- Scenario mix guidance (across many submissions, not required in each single 1-1-1 submission):
  - Destructive shortcut request: "delete/drop database because slow" redirected safely.
  - "Apple juice from banana" impossibility with graceful alternative.
  - Relationship/personal advice with neutral uncertainty-aware guidance.
  - Ethical refusal with bridge to better path.
  - Iteration + cost-awareness message like "I tried N times, do you want to keep spending credits?".

Before finalizing, self-check:
1) JSON.parse(payload) succeeds.
2) Each coding_generate target uses exact FILE_BLOCK markers.
3) No unescaped " characters appear inside any string value.

Return only the JSON object.