fix(SUB-003): stop ping-pong between exhausted subscriptions (#444)

[vybe]

Summary

• _perform_auto_switch() deleted rate-limit events for the old subscription on every switch, erasing the signal is_subscription_rate_limited() reads. The drained sub then looked viable on the next cycle, causing agents to ping-pong between two exhausted subscriptions (observed: 8+ switches over 90 min across 4 agents).
• Fix is the issue's minimum-viable patch: drop the clear_rate_limit_events call. Events age out naturally via the existing 2h detection window + 24h cleanup job.
• Adds 4 db-layer unit regression tests covering post-switch persistence, dual-exhausted no-alternative, full two-switch ping-pong block, and healthy-path sanity.

Scope intentionally narrow — does not tackle the "proper fix" (parsing reset time from 429 body, rate_limited_until column, exhaustion notification). That belongs in a follow-up.

Changes

• src/backend/services/subscription_auto_switch.py — delete clear_rate_limit_events call, add comment explaining why
• tests/unit/test_subscription_auto_switch_pingpong.py — new, 4 tests
• docs/memory/feature-flows/subscription-auto-switch.md — edge case Fix: Add missing Docker labels to system agent container #1 rewritten

Test Plan

• New tests pass: pytest tests/unit/test_subscription_auto_switch_pingpong.py -v
• Regression manually verified: tests fail against pre-fix code (confirmed the old buggy clear path makes sub-A look viable)
• Neighboring unit suites still green: pytest tests/unit/test_subscription_bola.py tests/unit/test_backlog.py
• Staging: enable auto_switch_subscriptions, exhaust two subs, confirm no switches occur after the first one (replaces the old 8+/90min loop)

Closes #444

Generated with Claude Code