ci: rebalance matrix groups based on actual timings
Split the two slowest groups based on measured wall-clock times:
- proxy-unit-b2 (was 7m15s, test_proxy_[s-z]*):
→ proxy-unit-b2: test_proxy_s*.py (server + setting_guardrails, ~3368 lines)
→ proxy-unit-b3: test_proxy_[t-z]*.py (utils + token_counter, ~3618 lines)
- proxy-unit-b3 (was 4m30s, test_[r-z]*):
→ proxy-unit-b4: test_[r-t]*.py (response_polling + search + skills + realtime, ~1988 lines)
→ proxy-unit-b5: test_[u-z]*.py (user_api_key_auth + zero_cost + update_spend + unit_tests, ~2394 lines)
proxy-unit-a2 (6m15s) will self-resolve once PR #21679 merges
(55 skip markers added to test_key_generate_prisma.py).
Matrix grows from 16 → 18 jobs; all groups expected ≤ 3-4m.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
c8ddbd90d1
