Frontend shell: the pristine theater drawn at / (closes #39)

[bgent19]

Closes #39.

make api now serves a hand-written SVG theater at /, drawn at the
instructor's authored coordinates, with /docs untouched. Zero dependencies:
no Node, no npm, no build step, no CDN, no web fonts. The assets live inside
the package and are mounted as static files, so pip install gets the frontend
with nothing extra to run — verified by building a wheel and confirming all
seven modules are inside it.

Structure

Seven modules with the ownership the spec fixes. api.js is the only one that
knows an endpoint URL; graph.js builds the DOM once while render.js only
ever sets geometry, text and classes — splitting the two makes build-once
structural, since appending from the render pass would be a visible
import-direction violation rather than an invisible line in a diff. Every
appearance value lives in one :root token block, so retuning for a projector
is an edit to one stylesheet.

Boot wiring is inline in index.html rather than a main.js: the spec fixes
the inventory at seven, and an eighth module would have been the implementer's
choice rather than the spec's.

graph.js freezes four solves against the pristine network, so no mark slides
when a lane is removed: the layered fallback, the directed-pair lookup (reverse
arcs have no id by construction), the sliding lane-anchor solve, and the
crossing solve. Casings are knockouts on the annotation tier only, so the
opening screen carries zero marks.

Verification

The frontend is verified from Python or not at all, and that is a rule rather
than an absence — no test asserts the SVG rendered; the picture is eyeballed,
forever. tests/test_web_assets.py covers the offline scan, the token subset
and the type floor. tests/test_scenarios.py gains the clearances and a 45°
crossing-angle floor — an angle floor rather than a crossing count, because
the count is a proven global floor that can never fire while the angles are a
windfall protected by nothing.

Each geometry check was confirmed able to go red rather than trusted on a green
run. The angle floor is the one that matters: nudging AIRHEAD-7 by -120,-120
holds every other check green and keeps the crossing count at exactly 8,
while dropping an angle to 42.5° — precisely the invisible regression a count
assertion would never catch.

The anchor solve is mirrored in tests/layout_geometry.py. The two were
compared coordinate-by-coordinate across all three datasets and agree to 6e-14.
That check needs a JS runtime, so it is run by hand rather than added to the
suite; layout_geometry.py documents how.

Two things worth a reviewer's judgement

• The spec's own numbers conflict. Frontend shell: the pristine theater drawn at / #39 names both a 74 anchor clearance and
  a 120 minimum lane length; those cannot both hold, since a lane needs 2×74 =
  148 to seat an anchor at all. A lane between 120 and 148 would pass the floor
  and then fail the anchor check. I derived the floor from the clearance and
  documented why. Nothing shipped moves — the shortest lane is 242.
• The ledger rows carry id/from/to. The spec defers "the ledger's content"
  to the panel ticket; I read those three as the row's identity rather than
  its content. Easy to trim.

Both review axes (standards + spec) were run and their findings fixed in the
second commit — the largest being that the opening screen was not actually
pristine, and that the anchor mirror diverged on infeasible lanes in a way that
would have shifted every anchor solved after one.

250 tests pass.