Python FastMCP servers have no route either: @mcp.tool is the largest measured registration idiom

Summary

#431 shipped a built-in registry of MCP tool-registration idioms covering TypeScript
and Go — the two languages its three vendor walks measured. The 30-server survey it
required before choosing that set is published at
docs/mcp-registration-idioms.md, and it
found one shape above the coverage line that #431 deliberately did not ship:

Python FastMCP is the largest single shape in the survey by both measures. redis/mcp-redis,
neo4j-contrib/mcp-neo4j, chroma-core/chroma-mcp and awslabs/mcp are all in the state the
MongoDB and Grafana servers were in before #431: detect returns is_agent_project: false.

Why it was deferred rather than skipped

It is a different extraction mechanism, and #393's lesson is that the mechanism is
reusable but each one needs its own adversarial probe list before it can claim anything.
Concretely, none of these hold for the lexical reader #431 shipped:

• The name defaults to the decorated function's name and is only sometimes a literal
  (@mcp.tool(name="…")).
• The input schema comes from the signature, so unlike TS/Go this route could carry
  real parameters — which changes what the SHIP-SCHEMA-* family sees.
• Provenance is an import binding (from mcp.server.fastmcp import FastMCP,
  from fastmcp import FastMCP), not a declared dependency, and the binding is what stops
  @anything.tool from reading as a registration.
• Python has a real AST in-tree, so a masking lexer would be the wrong tool: the existing
  inputs/_python_framework.py machinery is the right base.

The lexical reader had eight fail-open constructs in its first draft, all found by
writing the probe list. A Python idiom bolted onto #431's increment without its own list
would have shipped without that check.

Scope

• One idiom, py_fastmcp_decorator, in the existing built-in registry — same registry,
  same medium ceiling, same per-file completeness rule, same exclusion-ledger accounting
  for a name this reader cannot resolve.
• AST-based, following google_adk's binding-table approach rather than a text scan.
• Its own adversarial probe list, at minimum: a decorator on an alias
  (t = mcp.tool), a name built at runtime, a decorator applied by a factory,
  functools.wraps, a conditional registration, a re-bound module symbol, and
  @app.tool where app is not a FastMCP instance.

Acceptance

• detect on redis/mcp-redis and neo4j-contrib/mcp-neo4j returns
  is_agent_project: true with a named source.
• scan enumerates their tools, with parameters read from the signature.
• A decorator whose name is built at runtime is reported unenumerated and reaches the
  exclusion ledger, as in MCP servers whose tool surface exists only as code have no route: detect reports the official MongoDB and Grafana servers as not agent projects #431.
• The published survey table in docs/mcp-registration-idioms.md moves the Python row
  to "shipped".
• IDIOM_REGISTRY_VERSION bumps, and the trigger catalog's token list follows.

Filed from #431's survey; the measurement is the evidence the
roadmap non-goal clause asks for.