coli: the banner said GLM-5.2 over every model

[JustVugg]

coli chat, run, info, plan and tune all printed the same third banner line whatever you loaded:

GLM-5.2 · 744B MoE · int4 · streaming CPU

Start a chat with Inkling, Kimi K3, OLMoE or the new DeepSeek V4 engine and the banner still said GLM-5.2 · 744B. With four engines in the tree and a fifth arriving, that is no longer a cosmetic detail — it is the first thing the TUI tells you, and it was false four times out of five.

Where the wrongness actually lives

Not in banner(). In what banner() would have had to ask:

def model_arch(model):
    ...
    return "glm"          # every unrecognised model_type lands here

That default is right for its job. colibri.c is the general engine, so an unknown checkpoint should be offered to it rather than refused. It is simply the wrong answer to a different question — what did the user load — and the banner needs that one.

So model_banner_line() reads the raw model_type and leaves the dispatch in model_arch() untouched. No engine-selection behaviour changes.

What it prints now

GLM-5.2 · 744B MoE · 372 GB on disk
Inkling · 975B MoE · 469 GB on disk
Kimi K3 · 2.8T MoE · 1.1 TB on disk
DeepSeek V4 Flash · 284B MoE · 167 GB on disk
OLMoE · 7B MoE · 4.2 GB on disk
qwen3_moe · 48L x 128E MoE · 61 GB on disk        ← not in the roster

Three decisions worth stating, because each one is a place where it would have been easy to lie:

Parameter counts come from the README roster. Same numbers, one source; they cannot drift apart without someone noticing.

An unrecognised model_type is not forced into a name. It prints its own type and the geometry measured from its own config. Guessing "GLM-5.2" for a Qwen checkpoint is exactly the bug being fixed, and inventing a nicer-looking wrong answer is not an improvement over the old one.

A model with no config.json, or an unreadable path, keeps the original tagline. coli info banners before it validates the model directory, so this had to be safe on a bad path — there is a test for it.

Cost

Size is measured by stat-ing the shards. No safetensors headers are parsed: a banner runs before every command and may not cost a scan of a 400 GB checkpoint. The whole line is one config.json read plus one stat per shard.

For the other PRs that touch banner()

#825 and #670 both change banner("run") into banner("run", <cuda flag>), positionally. model= here is therefore keyword-only, so a second positional argument raises TypeError rather than being silently read as a path. test_model_is_keyword_only pins that.

Whichever of us lands second gets a trivial one-line conflict instead of a banner that tries to open True as a directory.

Tests

Eight new cases in tests/test_cli_output.py, run with the existing suite (16/16 green):

• every engine names itself
• the regression itself — a deepseek_v4 config must not produce a line containing GLM or 744B
• an unknown model_type states its own type and its measured geometry
• a config with no model_type says unknown model rather than inventing one
• no model, and an unreadable path, both keep the generic tagline without raising
• sizes do not round to 0 GB on a small checkpoint
• model= is keyword-only