feat(agent): harden memory fallback and MCP tool controls

[anfibiacreativa]

Summary

• Add deterministic fallback logic for project memory persistence when model tool-calls do not execute.
• Add MCP tool override enforcement and support for per-server request headers in agent-side flows.

Why

• Memory updates were not reliably persisted when the LLM skipped or missed write_project_memory.
• Teams need safer control over available tools and support for authenticated/private MCP servers.

What Changed

• Add server-side fallback path to synthesize and persist compact memory notes with stronger relevance gating.
• Add tool override loading and filtering so disabled tools are removed before model execution.
• Thread mcpServerHeaders through MCP tool listing and chat registration paths.
• Keep skill loading behavior aligned with config status to prevent draft content leakage.

Test Plan

• Run npm test in da-agent and verify memory/tool-override unit coverage.
• Manually send chat requests with and without mcpServerHeaders and verify tool availability.
• Confirm disabled tools from config are excluded in runtime tool registry.

Risks / Follow-ups

• Fallback memory writes may still need tuning to avoid noisy persistence on marginal relevance conversations.