Improve AI runtime validation/retries, reporting, and data ingestion robustness

[Rohan5commit]

Motivation

• Increase robustness of AI-enabled daily runs by validating configured inference endpoints, adding preflight predict checks and retries, and ensuring fallback behavior is reliable.
• Make reporting clearer and safer by preserving model-provided reasons, exposing AI runtime metadata and the performance period, and guarding daily email coverage.
• Harden data ingest and prediction client code paths to tolerate malformed responses and alternate endpoint shapes.

Description

• Workflow changes in .github/workflows/daily_trading_bot.yml: added disable_core_trading input, added retry steps for Core and AI runs, added AI runtime planning/validation steps that consider TRAINED_MODEL_INFERENCE_URL, added a preflight predict verification step, added environment tuning (TRAINED_MODEL_MAX_RETRIES, backoff, batch size), improved distilled-fallback and retry logic, and added an "Ensure Daily Email Coverage" guard and updated final status evaluation to include retry outcomes.
• Trained model client updates in trained_model_client.py: recognize alternate provider names, include TRAINED_MODEL_INFERENCE_URL in candidate envs, add alternate-prediction-url handling (retrying on 404), tolerate missing /health by short-circuiting readiness for some Cerebrium shapes, and improve retry/backoff/error reporting.
• AI prediction logic in llm_trader.py: add soft-cash neutral-breakout logic to allow converting strong neutral probabilities into directional signals when configured and there are no open positions.
• Reporting changes in email_notifier.py and main.py: preserve and display model-provided decision_reason for AI entries/exits instead of rewriting, include performance_period_start in Core and AI reports, and surface AI runtime/backend/model details in AI emails and in the ai_insight summary.
• Positions tracking in positions.py: added get_performance_period_start() to report the earliest entry date since last reset for a table.
• Data ingestion in ingest_prices.py: make Stooq fetch resilient by iterating candidate ticker shapes, skipping bad CSV lines, validating required columns, and logging attempt-level warnings instead of failing the whole loop.
• Router script quant_platform/scripts/plan_ai_runtime.py: include TRAINED_MODEL_INFERENCE_URL as an acceptable Cerebrium URL source.

Testing

• Ran the repository test suite with pytest -q; existing automated tests passed.
• Executed a local smoke run of the daily job and exercised AI runtime planning and the Stooq ingest paths to verify preflight predict, retry behavior, and artifact/report generation (manual smoke checks succeeded).

---

Codex Task