feat(cli): say when a generation went entirely to the reasoning channel

[inureyes]

Why

mlxcel generate suppresses the <think> channel by default. When a reasoning model's generation ends before the channel closes, every token is reasoning, the content channel is empty, and the CLI printed a blank line followed by its timing line. The model is working; the screen says nothing, and a blank screen is indistinguishable from a broken checkpoint or a broken patch.

That reading was one step away twice in one day. First on glm-4.1v-9b-thinking-4bit. Then on nvidia-nemotron-3-nano-30b-a3b-4bit, where the blank was the safety half of the granitemoehybrid RMS-norm A/B and attributing it to the arm under test would have inverted the verdict.

What changed

reasoning_stream::is_reasoning_only(generated_text, saw_visible_text, show_reasoning) names the condition: tokens were generated, none reached the content channel, and reasoning is suppressed. Both output paths consume it. The one-shot generate flow prints [All N generated tokens went to the reasoning channel; the content channel is empty. Re-run with --show-reasoning to see them.] in place of the blank. The chat REPL had the same hole on every turn and now prints the same line there. With --show-reasoning the predicate is false, because the text is already on screen.

scripts/ab_output_equality.sh is the output half of an A/B, run so that three recurring errors are not available. Sampling: a checkpoint's own generation_config.json can enable it with no flag from the caller, and then the files being compared are two samples rather than two implementations, so an untouched arm reads as a failure and a broken one can pass. The script passes --temp 0 to both arms and never takes it from the caller. The reasoning channel: comparing an empty content channel against an empty content channel passes while comparing no tokens at all, so both arms run with --show-reasoning. Attribution: a difference only belongs to the arm if the baseline reproduces itself, so the baseline runs twice as a control and a disagreement there reports INCONCLUSIVE (exit 2), pointing at the teacher-forced logit trace instead.

docs/benchmarks.md gains "The output half of an A/B" alongside the four conditions for judging a change that moves the numbers.

Validation

M1 Ultra, nvidia-nemotron-3-nano-30b-a3b-4bit at --temp 0. At -n 20 the generation truncates inside the channel and the new line prints; at -n 64 the channel closes, **Paris**. prints, and the line stays silent; non-thinking qwen2.5-7b-instruct-4bit stays silent. Three of three conditions behave, including both negatives.

The script was falsified against test doubles rather than only exercised: two identical binaries report EQUAL, a wrapper that alters the prompt reports DIFFERS, and a wrapper that alternates its prompt between invocations reports INCONCLUSIVE.

Five unit tests cover the predicate (unclosed thought with no answer, answer present, --show-reasoning on, nothing generated, whitespace-only content). cargo test -p mlxcel --lib reasoning_stream:: is 28 passing, cargo fmt --all -- --check is clean, and cargo clippy -p mlxcel --all-targets is clean. Workspace clippy is red on main for two unrelated lints in mlxcel-core, fixed separately in #1720.