MCP: expose the models catalog as gflow_list_models #883

Split out of #880 (gap 3).

gflow models --json returns the live image/video catalog — names, aliases, reference caps, durations, aspects, tiers — and MCP clients have no way to read it. They hardcode catalogs instead; the reporter in #880 ships a dated snapshot for exactly this reason.

Why the current exemption is wrong

tests/mcp/test_cli_parity.py::_MCP_EXEMPT says:

"models": "informational; models are enumerated in the generate tools' descriptions"

Prose in a tool description is not a catalog. It carries no ref_cap, no max_duration, no alias list, and nothing a client can parse. This is the second law's failure mode: a capability behind only one door, with nothing telling the agent at the other door that it exists.

Shape

cli_models.build_catalog() is already pure, Click-free and does no I/O — it derives everything from the Model / VideoModel enums plus the alias maps and per-model caps, so it cannot drift from what the generation commands actually accept. The tool is a wrapper over it.

gflow_list_models() -> dict   # build_catalog() verbatim

No arguments, no profile, no browser, no spend.

It must not touch the labs tRPC listing. projectInitialData 401s for migrated accounts (the #791 family) and project.createProject on that surface now answers 404 "Flow RPCs have been deprecated and disabled". Reading the static table keeps the tool correct on both hosts.

Done when

• gflow_list_models registered in src/gflow_cli/mcp/tools.py beside gflow_list_tools, added to __all__
• "models" moved from _MCP_EXEMPT to CLI_TO_MCP in tests/mcp/test_cli_parity.py
• docs/MCP.md lists the tool
• Run over a live serve and paste the result — the MCP twin is its own surface, so the Iron Law applies to it separately. Zero credits; tests/e2e/test_credits_mcp_e2e.py is the precedent at 3 assertions / 58s.