Draw the thesis lifecycle, and give crucible its own generated header

[HarperZ9]

The README explains the engine well and shows nothing. A reader has to assemble
the run from a feature list. This adds a header and a workflow diagram, and the
machinery that keeps both true.

The picture is generated from a spec

docs/art/crucible.art.json holds the words. Both SVGs are pure functions of
it, so tests/test_repo_art.py re-renders and compares bytes. Artwork that
stops matching its own text fails the suite instead of drifting quietly, which
is the normal fate of a picture in a README: nobody diffs it.

Byte stability across platforms comes from a fixed feTurbulence seed, a
random.Random(int) corona, 2-decimal coordinate formatting, and a renderer
that writes with newline="". The repository's existing * text=auto eol=lf
already covers checkout, so no .gitattributes change was needed.

The diagram is grounded in the code, not the highlights list

I read run_cmd._run_once and refine rather than working from the feature
bullets. Two edges carry the design.

The loop back from Cohesion to Measure is the refine step. Grading each claim
yields a margin, the harmonic mean of those margins is the cohesion, and a
round that is not cohesive names its weakest axis and measures that axis again
instead of reporting a thesis that only half held.

The edge that does not loop is verdict_for: a pure function of a deviation
and a tolerance, with no model in it. An axis that could not be measured reads
UNVERIFIABLE rather than holding.

New gate for a failure nothing was catching

test_the_tagline_stays_inside_its_rule. The header tagline is one unwrapped
line of 21px text under a rule that ends at x=700. A longer one runs on toward
the aperture and stops being readable, and the render still succeeds. The bound
counts characters, so it is a guardrail set at the widest tagline that has been
looked at on a rendered page, not a typographic measurement: it cannot tell
mmmm from iiii. Stated that way in the test so nobody reads it as exact.

Two gates carried over from the gather change also apply here:
test_no_card_note_loses_its_ending_to_the_wrapper catches the wrapper's
silent three-line truncation, and test_a_return_edge_stays_on_its_own_row
catches a backward edge routed straight through the cards between two rows.

Identity

The corona seeds from sha256 of the repository name, so crucible draws a
different aperture from flywheel and gather while keeping the same form.
test_each_repository_gets_its_own_drawing checks that rather than asserting
it in a doc.

Banner

.github/assets/zentropy-banner.png stays as the 1280x640 social-preview
source, uploaded through repository settings. It is no longer the README's
first image, so test_readiness now requires the two generated files in the
README instead of the banner. Its existence check is unchanged.

Checks

ruff, mypy and the full 372-test suite pass locally on Windows.

Honest null: the dark palette was verified on this same stylesheet on the live
gather README, not re-verified on a rendered crucible page yet. The forced-theme
case, where a signed-in reader pins GitHub to light while the OS reports dark,
is untested; GitHub's default is data-color-mode="auto", which follows the
same signal the SVG does.

🤖 Generated with Claude Code