Add AI inference validation & retries, improve reporting, and harden data/clients

[Rohan5commit]

Motivation

• Improve reliability of daily runs by validating and retrying AI and Core execution paths and ensuring end-to-end email coverage.
• Prefer and validate configured remote inference endpoints (Cerebrium / trained-model) and fall back cleanly to local distilled runtime.
• Simplify email rationale to surface model-provided reasons and surface AI runtime metadata in reports.
• Harden price ingestion and trained-model HTTP client to tolerate malformed CSVs, alternate endpoint shapes and transient errors.

Description

• Workflow: added disable_core_trading input, added core and AI retry steps, added a Validate Cerebrium primary configuration step, a Verify Cerebrium predict endpoint preflight that POSTs a sample payload, updated plan_ai_runtime usage to consider TRAINED_MODEL_INFERENCE_URL, and added an Ensure Daily Email Coverage guard and more granular failure reporting conditions in /.github/workflows/daily_trading_bot.yml.
• Email/reporting: updated email_notifier.py to stop rewriting model reasons and instead return model-provided decision_reason, include performance_period_start and detailed AI runtime metadata when subject_tag == 'AI', and adjust pipeline summary behavior.
• Ingestion: improved Stooq fetching in ingest_prices.py to iterate candidate tickers robustly, use on_bad_lines='skip', validate required columns, and log warnings rather than failing.
• LLM/Ai logic: in llm_trader.py added neutral-breakout soft-cash deployment logic that can convert neutral predictions into directional signals under controlled criteria and introduced guards using soft_cash_deploy_enabled and whether positions are open.
• Main/backtest and positions: added performance_period_start propagation to core and AI reports in main.py, suppressed rotation closes when the model signalled skipped reasons (e.g. all_neutral/no_action) and record a HOLD decision, and added get_performance_period_start helper to positions.py.
• Trained model client: in trained_model_client.py added support for additional provider name variants, improved retry/backoff logic including trying alternate prediction URL shapes on 404, made /health handling tolerant for providers that lack a /health route, and exposed additional settings like backoff and batch sizing.

Testing

• No automated test suite was executed as part of this change; changes were limited to CI workflow, email formatting, ingestion/client robustness, and runtime decision logic and should be validated in integration runs under the updated GitHub Actions workflow.

---

Codex Task