Trajectory capture, smarter conditionals, and vision-enriched knowledge bases

Every agent run can now capture its full tool-calling trail — including failed and rejected calls — as a trajectory artifact you can inspect from the execution's Artifacts view. It's off by default, with a global kill switch and a per-use-case toggle.

The CONDITIONAL node got a revamp: new is_empty/not_empty and array_contains/not_array_contains operators, LLM classification with token-usage reporting, and an expression editor with autocomplete, live validation, and duplicate-label detection. Execution logs now show the matched branch and evaluated result instead of a raw JSON blob.

Also in this release:

• Vision enrichment for knowledge bases: images, charts, and diagrams inside PDF, DOCX, and PPTX files are read by a vision model and merged into the extracted text
• File attachments in MCP tool calls: agents can pass uploaded or workflow-generated files straight to MCP tools by ID
• Audit logging extended to workflow import, export, and execution export, with filters and section grouping on the audit log page

We also fixed blackboard recall, which was silently erroring on every query, and hardened the expression evaluator against unsafe or invalid expressions being swallowed into a false branch.