guides: add Tested Models, the measured picture instead of a claim list

[webdevtodayjason]

What was missing

guides/models.mdx ended with a heading that said "Tested models include:" over a
list of model families that had never been benchmarked through AINode: Llama 3 405B,
Mixtral, Command R+, DeepSeek V3. Meanwhile the four runs we have actually measured
were sitting in bench/results/*.json in the product repo with no page pointing at
them and no numbers on the docs site at all.

So the docs asserted testing we had not done, and hid the testing we had.

What I decided

A new page, guides/tested-models.mdx, that defines its own bar. A model is on
it only if it was launched through AINode rather than a hand-written docker run,
somebody measured it on our hardware with engine-reported numbers, and the run is
committed as a JSON record a reader can open. Stated up front, so the page is
falsifiable.

Same table as the product README, same generator. The rows come from
scripts/render-bench-table.py in the product repo, spliced between
{/* BEGIN bench-table */} markers (JSX comments, since MDX rejects HTML ones). Each
row links to its JSON on GitHub. Regenerate with:

python3 scripts/render-bench-table.py \
    --readme ../ainode-docs/guides/tested-models.mdx \
    --begin '{/* BEGIN bench-table */}' --end '{/* END bench-table */}'

The gaps get a section, with a Warning callout saying they are real gaps rather
than omissions. Vision on Ornith 1.5 (launched text-only with
--limit-mm-per-prompt {"image":0,"video":0} after the multimodal warmup OOM-killed
the engine core). TP=4 through the current 0.5.x launch path. The Dell C4130 that
joined as a worker on 2026-09-11 with no AINode-launched model. The GLM-5.3-Flash
TP=2 pair that runs raw outside AINode. DeepSeek V4 and V4.1 Flash.

scripts/ainode-bench.py is described, not promised. It is not on main yet, so
the page says so in an Info callout, describes the intended shape from
bench/SCHEMA.md (one JSON per run in bench/results/, named
<stamp>-<model-slug>-<label>.json), and then shows the flags of
scripts/bench-serve.py, which is on main today and takes the same measurements.

A submit-a-result section with the one rule that matters: leave out a section you
did not measure. A missing key renders "not measured", which is the answer we want;
filling it with a plausible number is what gets a result rejected.

Fixed the claim that prompted the page. models.mdx now says "Model families
known to load and serve here include" and links to the new page in one sentence.
That list was never a list of tested models.

Nav: guides/tested-models goes right after guides/models in the Guides group.

Proof

mintlify broken-links  ->  success no broken links found
@mdx-js/mdx compile    ->  ok guides/tested-models.mdx, guides/models.mdx, guides/distributed.mdx
docs.json              ->  parses

The CLI would not run through npx on this machine (npm error ECOMPROMISED: Lock compromised during build:link), so mintlify was installed into a scratch directory
and run from there against this working tree.

🤖 Generated with Claude Code