docs: say what the engine is on the first screen, and where it came from

[inhuman]

Nothing in the engine changes here. What changes is whether a reader can tell
what the engine IS before deciding to close the tab.

What was wrong

The README opened with a mechanism and never answered "what is it". A reader
arriving cold had to assemble the answer out of six paragraphs, and the word
"skill" pointed them at the vendor feature of the same name — a folder of
instructions handed to a model — which is close to the opposite of what runs
here. The comparison section named two neighbours and left out the one people
ask about most.

And the provenance was wrong in a way that mattered: both files said the engine
was taken out of a working assistant and ran ~30 skills in production, which
reads as one story and is two.

The first screen

A sentence, then the numbers, then what it deliberately does not do:

A deterministic runtime for one turn of an LLM agent. The model decides
what; the code decides what runs next. A restriction written into a prompt
is a request; a branch that does not exist cannot be taken.

Three sections follow it. What it is — what a skill, the engine and the
application each own, plus a paragraph on the name for readers who arrive with
the other meaning of the word. In 30 seconds — the same turn end to end:
the YAML, six lines of Go, and a real trace from examples/simple-llm-app
rather than a claim about one. Architecture — the picture below.

The picture

The diagram was ASCII art: it wrapped on a phone, read as a code sample, and
could show which side of a boundary a thing was on only by indentation. It is
now an SVG in a <picture> that follows the reader's theme, with the ownership
boundary drawn — the application's frame around everything, the engine as the
one block inside it, five named ports leading back out, and the return path for
what comes back.

Four files, because a picture with words in it has a language. They are
generated by assets/architecture.py, and two guards keep them honest:

• readme_test.go compares them with each other — two themes of one
  language may differ only in their palette, and the two languages must draw
  the same number of boxes, lines and captions. It found a defect while being
  written: Runner carried two lines of caption in Russian against one in
  English;
• four files edited by hand in the same wrong way would still agree with each
  other, so CI regenerates them and fails on any difference.

The link check now follows src and srcset too: a theme-specific image is
HTML rather than markdown, and a missing one shows as a broken icon to exactly
half the readers — the half whose theme nobody tested in.

Where it fits

"What this is not" became a table where every neighbour is named with what it
does better, then the architectural reason this is not that: LangGraph,
Temporal, n8n and Airflow, BPMN engines, and MCP. MCP was missing entirely and
is the comparison asked for most; the answer is that it is orthogonal — MCP
says how a call is made, this says which call happens and when — and saying so
is more useful than a list of differences.

Where it came from

Status now carries two facts as two paragraphs. It was written for a home agent
on a single 8 GB GPU, where a generation costs minutes somebody sits
through and the token budget is the card rather than an invoice — which is the
constraint that gave the format its shape. It was adopted by an assistant in
production afterwards, and that is where the catalogue and the measurement come
from.

The order matters beyond accuracy: told the other way round, the same design
looks like an architecture chosen for its own sake.

Also

Everything below the documentation map is now marked as reference, with an
index, because the file is long and a first reader should be able to stop.

Stale numbers, fixed in passing: library v0.5.x → v0.8.x, format 2.2.2 →
2.2.4, coverage 89.3/95.1 → 90.1/95.2, and the linter's rule count 29 →
30 in both READMEs and both quickstarts.

Checks

Vet, Test, Race, govulncheck and Examples as usual, plus the new Diagrams step.
That one was verified by reintroducing the defect: an SVG edited by hand and
staged makes the step fail, and regenerating makes it pass.