Pin the spec to LF, and give the outcome gate a control that can fail it

[HarperZ9]

Three changes to the artwork gate. No drawing changes.

The spec was not pinned

.gitattributes pinned docs/art/*.svg to LF but left docs/art/*.art.json
unpinned. Measured on this checkout:

	bytes	endings
index.art.json, before	2354	CRLF
index.art.json, after	2318	LF

2318 is what render_repo_art.py actually wrote. The gate's receipt records a
byte count and a SHA-256 read from the checked-out file, so unpinned those two
numbers depend on who cloned the repository.

What the pin does not do. It does not make the comparison work. Both sides
are read through Path.read_text, which applies universal-newline translation,
so the comparison passes on a CRLF checkout with or without the pin. Only the
receipt's size and hash fields are byte-sensitive. The comment in
.gitattributes now says that narrower thing instead of the broader claim it
carried before.

The budget had the geometry copied into it

_outcome_budgets computed the box width from literals: (960 - 44 * 2 - 26 * (count - 1)) / count. Move a margin in repo_flow and the gate keeps passing
while measuring a box that no longer exists. It reads FLOW.W, FLOW.PAD and
FLOW.GAP now, so the two move together.

The check had no control

The overflow logic is extracted into _outcomes_that_overflow so
test_that_check_can_actually_fail can feed it two outcomes it must reject.
Without that, a green run proved only that the check ran.

The module docstring also claimed the test compares bytes. It compares the
rendered text with line endings normalized away. Corrected.

13 art tests pass. Both budgets still count characters rather than measure
glyphs, so neither can tell mmmm from iiii: a guardrail, not a typographic
fact.

🤖 Generated with Claude Code