feat(tools): let MCP_TOOLS register a subset of tool groups

[wromansky]

Why

All 81 tools are registered unconditionally today. That is a lot of schema for one agent to carry, and it is awkward for two kinds of deployment:

• Agents driven by small local models. Tool-selection accuracy falls off as the surface grows, and every schema is spent from a modest context window. This is the common shape for self-hosters running a local model beside their own wger.
• Single-purpose agents. A food-logging assistant has no business holding 24 routine-authoring tools.

The handoff notes list several open items, but not this one — it surfaced while wiring the server up to a self-hosted instance for exactly that small-local-model case.

What

MCP_TOOLS takes a comma-separated list of group names, using the same CSV convention as ALLOWED_HOSTS. Empty keeps today's behaviour, so the change is backwards compatible.

MCP_TOOLS=nutrition,off,profile     # a food-logging agent

Valid names are the module names: profile, routines, workout_logs, body_weight, measurements, equipment, nutrition, exercises, analytics, off.

Three details worth a look during review:

• _REGISTRARS becomes a name-keyed dict and TOOL_GROUPS is derived from it, so the documented list cannot drift from the code.
• Registration iterates that dict rather than the env var. Tool order stays stable however the variable is written, and a repeated name is harmless rather than a duplicate registration.
• An unknown group name raises at startup, naming both the typo and the valid groups. Silently registering nothing seemed like the failure most likely to cost a self-hoster an evening.

Tests

8 new tests in tests/test_tool_selection.py: the default surface, a subset, duplicates, env ordering, the unknown-name error, CSV parsing, and that every name in TOOL_GROUPS is independently valid — so the documented list stays honest as groups are added.

Full suite passes (79) and ruff check is clean. MCP_TOOLS was added to the conftest cleared-vars list so a stray value in the environment cannot skew other tests.

Docs

README.md gains a short section under Tools, and .env.example a commented block next to the other optional settings.