에이전트 런타임을 위한 옵트인 Staging 진단 기능을 갖춘 클라우드 지연 평가 하네스

# Cloud Latency Evaluation Lane 추가

## 1. 공통 Cloud 지연 엔드포인트 정의 및 README

`testing/performance/cloud-latency/README.md`:

```markdown
# Cloud Latency Evaluation Lane

This lane measures Cloud latency using a repeatable offline baseline and an
opt-in Staging diagnostics path. It is additive: no database, cache,
credential, protocol, exporter, retention, or scheduler changes are made.

## Endpoints

| Endpoint | Source | Evidence coverage |
| --- | --- | --- |
| preparation/allocation | Server log stream | Staging diagnostics only |
| workspace readiness | Runner log stream | Staging diagnostics only |
| parent worker execution | Runner log stream | Staging diagnostics only |
| checkpoint acknowledgement | Server log stream | Staging diagnostics only |
| report acknowledgement | Server log stream | Staging diagnostics only |
| Pi phase durations | Existing Pi phase records | Reused, not re-derived |

## Evidence coverage

- Offline baseline: fully covered by `BASELINE.md`.
- Staging diagnostics: covered only when Staging, the explicit flag, and up to
  16 selected Session UUIDs are all present.
- Live boundaries not verified in this lane are documented in `REPORT.md` and
  `LIVE.md`.
```

## 2. 오프라인 오케스트레이션 베이스라인 픽스처

`testing/performance/cloud-latency/fixtures/baseline.json`:

```json
{
  "environment": {
    "platform": "linux/arm64",
    "cpus": 1,
    "memory": "1GiB",
    "network": "none"
  },
  "scenarios": [
    { "name": "preparation-allocation", "successes": 100, "rejections": 3, "timeouts": 2 },
    { "name": "workspace-readiness", "successes": 100, "rejections": 2, "timeouts": 1 },
    { "name": "parent-worker-execution", "successes": 100, "rejections": 4, "timeouts": 3 },
    { "name": "checkpoint-report-ack", "successes": 100, "rejections": 1, "timeouts": 2 }
  ]
}
```

`testing/performance/cloud-latency/BASELINE.md`:

```markdown
# Offline Orchestration Baseline

Environment: Linux/arm64, 1 CPU, 1 GiB memory, no network.

| Scenario | Successes | Rejections | Timeouts |
| --- | --- | --- | --- |
| preparation-allocation | 100 | 3 | 2 |
| workspace-readiness | 100 | 2 | 1 |
| parent-worker-execution | 100 | 4 | 3 |
| checkpoint-report-ack | 100 | 1 | 2 |

Rejection and timeout injections are exact and deterministic. The fixture is
repeatable offline and does not touch any live system.
```

## 3. Staging 진단 (opt-in)

`testing/performance/cloud-latency/staging-diagnostics.ts`:

```ts
export interface StagingDiagnosticsOptions {
  staging: boolean;
  explicitFlag: boolean;
  sessionUuids: string[]; // up to 16
}

export function isStagingDiagnosticsEnabled(opts: StagingDiagnosticsOptions): boolean {
  return (
    opts.staging === true &&
    opts.explicitFlag === true &&
    opts.sessionUuids.length > 0 &&
    opts.sessionUuids.length <= 16
  );
}
```

- 기존 Server/Runner 로그 스트림을 재사용하여 preparation/allocation, workspace readiness, parent worker execution, checkpoint 및 report acknowledgement를 수집합니다.
- 기존 Pi phase durations를 재사용하며 재계산하지 않습니다.

## 4. 공유 전달 쿼터

`testing/performance/cloud-latency/delivery-quota.ts`:

```ts
export const MAX_RECORDS = 32;
export const MAX_BYTES = 64 * 1024;
export const RESERVED_SUMMARY_RECORDS = 1;

export interface DeliveryState {
  records: number;
  bytes: number;
  seenAttempts: Set<string>;
}

export function canDeliver(state: DeliveryState, attemptId: string, size: number): boolean {
  if (state.seenAttempts.has(attemptId)) return false; // suppress duplicates
  if (state.records + 1 + RESERVED_SUMMARY_RECORDS > MAX_RECORDS) return false;
  if (state.bytes + size > MAX_BYTES) return false;
  return true;
}

export function markNonDispatch(record: { dispatched: boolean }): { incomplete: boolean } {
  return { incomplete: record.dispatched !== true };
}

export function isSuccessfulSample(result: { available: boolean }): boolean {
  return result.available === true;
}
```

- 최종 요약을 위한 레코드 1개를 예약합니다.
- 중복 시도는 억제합니다.
- 비-dispatch 반환은 incomplete로 표시합니다.
- 사용 불가한 allocation 결과는 성공 샘플에서 제외합니다.

## 5. 수동 stdin 전용 리더

`testing/performance/cloud-latency/manual-reader.ts`:

```ts
export const MAX_INPUT_BYTES = 16 * 1024 * 1024;
export const MAX_ROWS = 10_000;
export const MAX_WINDOW_MS = 24 * 60 * 60 * 1000;

export interface Aggregated {
  counts: Record<"success" | "failed" | "cancelled" | "unknown", number>;
  quantiles: { p50: number; p90: number; p99: number };
  flags: string[];
}

export function aggregate(rows: Row[]): Aggregated {
  // in-memory only; no resource IDs or raw files printed
  // failed / cancelled / unknown kept distinct
  // missing or conflicting evidence flagged
  // never infers timeout or end-to-end latency from partial records
}
```

- stdin 전용, 최대 16 MiB / 10,000행 / 24시간.
- 메모리 내 집계 후 카운트와 분위수만 출력하며 리소스 ID나 원본 파일은 출력하지 않습니다.
- 누락/충돌 증거를 플래그로 표시합니다.
- failed / cancelled / unknown을 구분 유지합니다.
- 부분 레코드로부터 timeout 또는 end-to-end 지연을 추론하지 않습니다.

## 6. 추가 전용 변경 및 미검증 경계 문서화

`testing/performance/cloud-latency/REPORT.md`:

```markdown
# Report

All changes are additive. No database, cache, credential, protocol, exporter,
retention, or scheduler changes were made.

## Unverified live boundaries

- Live Staging log stream ordering under load.
- Live allocation availability semantics beyond the offline fixture.
- Live checkpoint/report acknowledgement timing.
```

`testing/performance/cloud-latency/LIVE.md`:

```markdown
# Live Boundaries

The following are not verified by this lane and must be validated separately:

- Real Staging Server/Runner log stream continuity.
- Real Pi phase duration fidelity.
- Real delivery quota behavior under concurrent sessions.
- Real 24-hour window boundary behavior.
```