Refactor and harden front

[mborne]

• create services for agent and db
• closes Use a JSON file for MCP configuration #56
• closes create_map - turn GET into POST for cql_filter #57
• closes create_map - ensure that the LLM doesn't try to render multiple layers #58
• closes sanitize_html=True #54 (using gr.HTML for the map)

[esgn]

09c8528 => translate tool create_map description in french to be aligned with geocontext tools

[esgn]

0a45171 => /health/db was calling get_database() as if it returned a database instance, but get_database is an async context manager (@asynccontextmanager).
As a result, the endpoint received an _AsyncGeneratorContextManager object and then failed on await db.is_healthy() with AttributeError.

This was fixed by running the health check inside an async context (async with get_database() as db) via a dedicated helper (is_database_healthy()), so the endpoint now handles DB initialization correctly and returns proper HTTP statuses instead of crashing.

d675ca9 => Added explicit Redis client cleanup in get_database() by wrapping the Redis branch in try/finally and calling await redis_client.aclose(). This prevents connection leaks for short-lived contexts (e.g., /health/db) while preserving normal runtime behavior for the long-lived agent context.

[esgn]

8ece2f6 => This commit adds a minimal filter in to_gradio_message() to prevent raw map markup from being shown in chatbot text responses.
For ai messages, the content is now sanitized into cleaned_text by removing <ol-simple-map ...></ol-simple-map> snippets. If nothing remains after cleanup, the message is skipped; otherwise, only the cleaned text is displayed.

Scope is intentionally simple and stupid for now:

• No change to create_map tool behavior
• No shared state or message tracking
• No architecture changes

Regression coverage was added for:

• AI message containing only map markup -> hidden
• AI message containing text + map markup -> text is kept only

[esgn]

c09c0ab => This commit hardens the bot() async generator guard paths for invalid/empty submissions.
It now short-circuits when history is empty, the last item is not a user message, or user content is blank, and emits yield history before returning so the UI keeps the current conversation state (no accidental replay and no history reset on empty input).