Drop the internal daily cap on YouTube analytics jobs

[paulocastellano]

Problem

Nightwatch #18 (MaxAttemptsExceededException), 528 times in 24 hours, for CollectPublicationMetrics. The job fails at pickup in under a millisecond, which Worker::markJobAsFailedIfAlreadyExceedsMaxAttempts only does once retryUntil (the end of the day after the observation date) has passed. The jobs were not slow: they were never let through in time.

AnalyticsRateLimits gave YouTube a project-wide Limit::perDay(3000) (youtube:project), shared by metrics and discovery jobs. Once it runs out, RateLimited releases every YouTube job until the 24-hour window resets. The window is fixed and anchored at its first hit, so every blocked job comes back at the same instant, and the ones past retryUntil fail together:

• All 23 failures on the prod clone are YouTube metrics jobs observed on 07/10 and failed at the same second (09/10 13:19:46). Nightwatch's last occurrence is 10/10 13:21: the same minute, a day later.
• The large channels whose discovery was held behind the same key first appeared at 13:19:30 on 08/10 and 13:19:39 on 09/10.
• 07/10 (the feat: independent social posts, workspace analytics and media editing #374 launch backfill of channels with 450–506 videos each) had 2,798 YouTube reads. 730 videos of four channels were never measured; none had failed, they never ran.

The cap was far below what it was protecting. It assumed each job spent up to 2 Data API units and shared the 10,000-unit quota with uploads. In reality:

• A metrics job reads videos.list in batches of 50 ids, one unit for 50 videos (videos.list), plus one YouTube Analytics API report.
• videos.insert has its own bucket of 100 calls a day (quota costs, getting started).
• The project console shows 1,637 of 10,000 Data API units used, and 1,575 of 100,000 Analytics API queries. No quota went above 90% in 7 days.
• The cap also counted every job that passed the middleware, including discovery, duplicates and jobs that end without reading.

Fix

Remove the YouTube shared daily cap. The per-channel limit (30 jobs a minute) stays. A quotaExceeded, rateLimitExceeded or 429 from Google is already classified as rate_limited by AnalyticsCollectionException::fromResponse and retried; if that happens, the project asks Google for more quota.

Tests

• PublicationMetricsJobsTest: YouTube is limited only per channel (fails with the old cap).
• Full suite in parallel (PostgreSQL): 9577 passed. One unrelated test, RemoteMediaImporterTest "a transfer that runs…", failed once in the parallel run; it compares temp file names, and it passed 4 of 4 times on its own, both with and without this change.