A bare backend name was read as a model name, and the CLI's failures …

[Shashankss1205]

…said nothing

Two defects that hid each other. Both found by installing 0.1.2 from PyPI into a clean venv and driving it against a real claude v2.1.220.

split_spec only consulted BACKENDS when the spec contained a slash. A bare backend name fell through to "assume the default backend, keep the whole string as the model", so:

split_spec("claude-cli") -> ("claude-cli", "claude-cli")
split_spec("mock")       -> ("claude-cli", "mock")

--model claude-cli therefore shelled out to claude -p --model claude-cli and was refused by the CLI on every call (3/3), while grapharc models --check went on reporting the backend usable and grapharc models claude-cli printed model: claude-cli without complaint — two commands whose job is to say whether a spec will work, both saying yes about one that never did.

--model mock was the worse half: it named the paid subscription backend and spawned the real binary, so the double models --check describes as "scripted test double; never reaches a provider" reached for one. It happened to fail before billing only because mock is not a model name.

This is the same "silently folded into a model name … fails much later with a confusing error" failure split_spec already refuses for a mistyped backend with a slash; it just could not see the case without one.

A bare backend name now resolves to that backend. claude-cli takes its own default model and mock takes the scripted double (which ignores the model segment entirely). openrouter, openai and ollama front catalogues rather than a model, so a bare name there is refused with a spelling that works rather than a guess about what to bill someone for. Slash forms and bare model names are untouched.

The gateway read the wrong stream. claude -p fails with a non-zero exit, an empty stderr, and its whole explanation in the JSON envelope on stdout:

{"is_error": true, "result": "There's an issue with the selected model
 (claude-cli). It may not exist or you may not have access to it."}

_invoke_cli reported proc.stderr, so the error was claude -p exited 1: — a sentence that stops at the colon. That is how the bug above presented itself: as no message at all. The one string that would have diagnosed it in seconds was captured, held in proc.stdout, and discarded. Note the code already parses this shape a few lines below, but only on the returncode == 0 path, and the CLI sets is_error and exits non-zero, so it took the branch that ignores stdout.

stdout is read first now, falling back to stderr when it is not the documented envelope. The recovered text also feeds _cli_failure, which classifies transient-vs-deterministic and was previously classifying from "".

Tests: every new test was confirmed to fail against the old code and pass against the new — including the mock one, which asserts no subprocess is created rather than just checking the returned type, since the type was what was wrong and a refactor could fix the type and still shell out. BARE_BACKEND_MODEL["claude-cli"] is a second copy of the model class's default so the registry need not import a backend to split a string, and a test pins the two together the way CI already pins __version__ to the packaged version.

Verified: grapharc demo stage1 --model claude-cli now completes (8 nodes, target_met, exit 0) where it previously died with an empty error; --model mock resolves to the double with no subprocess; --model openrouter exits 2 with an example. Full suite green on 3.12 and 3.13; ruff clean.