Waive stale MAE gate for direct-net repair

[sheawinkler]

Summary

• Record active direct-net MAE/health in trainer status artifacts.
• Preserve the global-MAE degradation check, but waive it only when the candidate clearly repairs a pathological active direct-net head, corpus coverage expanded materially, candidate direct-net health is OK, and Brier is not degraded.
• Surface the decision as waived_issues=["global_mae_degraded_direct_net_repair"] plus explicit gate booleans instead of hiding the MAE degradation.

Evidence

• Live candidate 20260519T130158Z-4ebc8ed7: active direct-net MAE 0.36497176542453535; candidate direct-net MAE 0.005188941983922394; candidate health OK; corpus expanded 49061 -> 183757 shadow rows.
• Patched gate replay on that artifact returns ok=true, issues=[], waived_issues=["global_mae_degraded_direct_net_repair"].

Verification

• python3 -m py_compile src/api/predictive_trainer.py tests/sidecar/test_predictive_trainer.py
• uv run --with ruff ruff --isolated check src/api/predictive_trainer.py tests/sidecar/test_predictive_trainer.py
• PYTHONPATH=/Users/sheawinkler/Documents/Projects/fastapi_sidecar uv run --with pytest --with pytest-asyncio pytest -p pytest_asyncio tests/sidecar/test_predictive_trainer.py -q -> 65 passed

Runtime

• Python-only sidecar change; no Rust rebuild gate.

[sheawinkler]

Review evidence packet:

• Current blocker this PR addresses: latest sidecar trainer run 20260519T130158Z-4ebc8ed7 is gated_off on gate:global_mae_degraded, while active model is stale and direct-net pathological.
• Active direct-net MAE: 0.36497176542453535.
• Candidate direct-net MAE: 0.005188941983922394.
• Candidate direct-net health: ok=true.
• Shadow corpus expansion: active 49061 shadow rows -> candidate 183757 shadow rows.
• Candidate Brier: 0.15491637157440985; active Brier 0.13098749826096534; below configured degradation factor 1.25, so not p_positive_brier_degraded.
• Patched gate replay on the actual artifact returns ok=true, issues=[], waived_issues=["global_mae_degraded_direct_net_repair"]. The global MAE degradation is still reported, not hidden.

Verification run locally on branch fix/direct-net-repair-promotion-gate:

python3 -m py_compile src/api/predictive_trainer.py tests/sidecar/test_predictive_trainer.py
uv run --with ruff ruff --isolated check src/api/predictive_trainer.py tests/sidecar/test_predictive_trainer.py
PYTHONPATH=/Users/sheawinkler/Documents/Projects/fastapi_sidecar uv run --with pytest --with pytest-asyncio pytest -p pytest_asyncio tests/sidecar/test_predictive_trainer.py -q

Result: 65 passed.

Runtime note: Python-only sidecar change. No Rust rebuild gate.