feat(art): draw what a wiring edge settles and what it only repeats

[HarperZ9]

The README already says an edge is declared and never probed. Nothing drew it,
so a reader had to carry that sentence through every key plexus discover
returns and decide for themselves which ones it applies to.

docs/art/edge-evidence.svg draws the six keys an edge comes back with and says,
per key, what settled it.

key
producer	COMPUTED	the organ whose emitted port satisfied the name
consumer	COMPUTED	the organ that named the capability as an input
capability	COMPUTED	matched by equality or through consumable_as
self_loop	COMPUTED	true when both ends are the same organ
via	COPIED	the producer's own pointer, never resolved
evidence	CONSTANT	always declared

The hot mark sits on via. It is the field that looks like a citation and is the
one plexus never follows, which is exactly the claim mesh.py makes about itself
in its own docstring.

Held to the code, not to the picture

The existing gates settle whether the card fits its columns. Whether it is TRUE
of mesh.py and cli.py is a different question, and tests/test_repo_art.py
now drives a real mesh to answer it:

• the drawn keys are compared to the keys _mesh_json actually returns, in
  order, so a key added or renamed on either side fails
• every edge in the live mesh is checked to carry evidence == "declared"
• self_loop is checked to follow from the two ends, and checked to survive
  into the output rather than being dropped, because the note claims both
• via is proven copied by building a manifest whose module points at a path
  that is not in this repository, then watching the edge form with that string
  intact and evidence still declared
• the footnote says the edge holds a second pointer, the consumer's own module,
  which discover keeps and never serializes. Both halves are checked, because
  a field quietly added to the receipt would make the footnote wrong.

What came in with it

The card class: repo_card.py, check_repo_card.py, and the flow budget
helpers split out into check_repo_flow.py so no one file outgrows what a
reader can hold at once. Three card gates are registered, plus a fourth:

art.card_alt_reaches_the_readme binds the README's alt attribute to the alt in
the spec. GitHub draws the card as an <img>, and an <img> hides whatever
description the SVG carries inside it, so the README alt is the whole of what a
reader who cannot see the card gets. It is proven twice: the built-in control
card carries a description that appears in no README, and a one-word mutation of
this repository's own README was run through the gate, caught, and restored
byte-identical.

Measured

Against the 916px rule, with the canon face loaded: widest drawn text 913.4,
zero of thirty-two elements over. Rendered as a standalone document, where no
webfont loads and the fallback face is narrower, which is what a GitHub reader
gets from an <img>: 909.0, zero over.

80 tests pass. Sixteen art gates green.

Not done here

The three flow drawings still carry short README alts that do not match their
longer spec alts. On GitHub those are weaker than they should be. Rewriting them
across the fleet is its own change.

The renderer is a copy of the same files that live in eleven sibling
repositories rather than a shared package. That is a known duplication, carried
so each repository stays standalone.

🤖 Generated with Claude Code