fix(scheduler): preserve pre-hydration account metadata

[lyy0709]

Summary

Scheduler snapshots store a slim account projection in Redis, then apply scheduling filters before the selected account is hydrated from the database. Several fields used by those pre-hydration filters were omitted, so cache hits could behave differently from database-backed scheduling.

This change:

• preserves the minimal credential and extra fields consumed before hydration
• keeps access and refresh tokens out of scheduler metadata
• versions the metadata namespace as sched:meta:v2: so existing narrow snapshots cannot remain silently valid
• removes matching legacy metadata when accounts are written, deleted, or rejected as unencodable

The affected behavior includes per-account scheduling thresholds, Anthropic/Grok native usage windows, Grok free-tier/team gates, and OpenAI privacy/passthrough/compact routing.

This is a separate instance of the projection-loss pattern discussed in #5607; it does not duplicate #5609, which addresses API-key authorization snapshots.

Verification

• cd backend && make test-unit
• targeted repository race tests for projection and metadata-version cleanup
• relevant scheduling, compact, privacy, and Grok service tests
• cd backend && golangci-lint run --timeout=15m ./internal/repository/...
• git diff --check upstream/main...HEAD
• exact-head fork CI passed all test, integration, lint, frontend, and shell jobs: https://github.com/lyy0709/sub2api-upstream/actions/runs/31834418363

The exact-head Security Scan currently reports only two vulnerabilities already present on upstream main. They are handled independently by #5638 (nanoid) and #5639 (Go 1.26.6), so this scheduler PR intentionally does not duplicate those dependency changes.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.