Add a manual bounded investigation workbench using the read-only tool API

[VibeSWE]

The server already exposes authenticated dataset discovery, schema, SQL and trace inspection through its tool API. Add a manual workbench to select a supported tool, inspect arguments, explicitly run it and review evidence. Keep the existing Agents page focused on observed telemetry.

Use /api/prism/v1/llm/tools/list for capabilities/schemas and /api/prism/v1/llm/tools/call under the current session. Start with a small supported read-only subset; expensive reads remain explicit user actions. Server authorization stays authoritative.

Acceptance criteria

• Validate arguments against supported catalog schemas and explain unavailable capabilities.
• Bound time ranges, result sizes, concurrency and duration; support cancellation and document whether it stops execution or only discards the response.
• Display isError: true failures even with HTTP 200. Distinguish authentication, transport, malformed-result and tool errors.
• Retain tool name, arguments, dataset/time scope and source SQL where applicable alongside results, so evidence can be inspected or reproduced.
• Safely render ingested text as untrusted data; embedded instructions never authorize another tool call.
• Tests cover restricted sessions, failure envelopes, cancellation and stale responses.

Provider keys, autonomous execution, persisted model conversations and an AI orchestration service require separate scope.

Suggested order: a later bounded experiment; avoid treating existing tool endpoints as a complete assistant service.