fix(tradingagents): preserve configured timeouts and expose provider errors

[TNT-Likely]

Background

TradingAgents configurations saved by the UI include auto_trigger, which the agent constructor did not accept. The execution entry point caught the resulting TypeError and silently created a default agent, resetting a configured 600-second model timeout to 300 seconds. Provider failures such as GLM HTTP 400/code 1301 were also replaced with a generic request error, hiding the actionable reason.

Changes

• Accept the linked-trigger configuration without resetting execution settings, read it from the actual AgentConfig.config ORM field, and log effective timeout and model settings.
• Remove PanWatch's output token cap from seeded settings, saved UI configurations, and model client parameters. Retain the legacy constructor argument as an ignored compatibility field.
• Persist bounded, redacted HTTP status and provider code/message diagnostics for manual and scheduled runs; preserve those details in progress responses.
• Restore the latest failed run when reopening the deep-analysis dialog, wrap long error text, and ignore late recovery responses after switching stocks.

Validation

• .venv/bin/python -m pytest tests/test_ai_errors.py tests/test_tradingagents_agent.py tests/test_tradingagents_auto_trigger.py tests/test_tradingagents_progress_recovery.py tests/test_tradingagents_v050_compat.py tests/test_tradingagents_trigger_idempotency.py -q — 86 passed.
• pnpm --dir frontend exec vitest run tests/analysis/deep-analysis-modal.test.tsx tests/settings/deep-analysis-timeouts.test.tsx — 9 passed.
• Frontend check:ui, check:i18n, check:market-colors, and build — passed.
• git diff --check origin/main...HEAD — passed.
• A live GLM-5.3-flash analysis using existing local settings completed all 17 model calls in 45 minutes 22 seconds. Individual calls lasting 386 and 398 seconds completed successfully with the configured 600-second timeout.
• Checked desktop and 390 px mobile layouts in light/dark themes, scrolling, Escape dismissal, and reopening the live task and saved report.
• Verified the original HTTP 400/code 1301 response using the installed LangChain exception type and failure-restoration regression tests.

Boundaries and risks

• No database migration or model reasoning-depth changes.
• Provider moderation remains in effect. The live run did not reproduce code 1301; its diagnostic formatting and failure view are covered by the supplied response and regression tests.
• Removing the host output cap uses provider defaults and may increase output length and latency. Existing overall and request timeout settings remain available.
• Local proposal and QA files under .docs/ are excluded from this PR.

Follow-up

None required for this release.