feat(assistant): add trusted results and actionable research

[TNT-Likely]

Background

Assistant answers were difficult to verify after a run completed: factual claims were mixed with analysis, source timing was not durable, execution traces disappeared after refresh, and useful conclusions still required manual navigation or repeated prompts. This made long-running research harder to trust and act on.

Changes

• Added a versioned structured assistant result with evidence, deterministic facts, inferences, risks, missing data, and next actions.
• Persisted tool result data, source metadata, observation times, structured results, and final-message links through migration 128.
• Added deterministic freshness, missing-field, repeated-quote, and quote-versus-close checks while keeping the optional model pass limited to organizing analysis.
• Restored structured results and user-visible execution traces when historical conversations are reopened, including a safe summary fallback for legacy tool-backed answers.
• Added result cards with expandable evidence, validated internal actions, price-alert proposals, and direct K-line navigation.
• Added portfolio deep-link handling for validated market and symbol parameters.
• Split the trace into user-facing research progress and expandable developer details.
• Enforced fail-closed tool restrictions for explicit action requests and persisted source timing from the runtime through the API.

Validation

• .venv/bin/python -m pytest tests/test_assistant*.py packages/pan-agent-runtime/tests -q — 143 passed.
• pnpm test -- --run in frontend/ — 89 passed.
• pnpm build in frontend/ — production build succeeded.
• git diff --check and git diff --cached --check — passed.

Boundaries and risks

• Legacy answers created before field-level tool data was persisted are restored from their durable tool summary and explicitly marked as lacking field-level history.
• The local frontend environment provides Node.js 20.20.2 while the repository requests Node.js 24.14.0. Tests and the production build passed, with pnpm reporting the engine mismatch.
• Freshness thresholds and material price-difference thresholds are deterministic product defaults and may need tuning with production usage.

Follow-up

• None required for the R1 scope.