fix(layout): bind the locals these three paths read on every route

[lmeyerov]

Fourth of the pyright-ratchet cleanup series (#2092 gate, #2093 layout, #2094 outliers, #2095 plotter). Ratchet 262 → 255.

Each of these reads a name that an earlier branch may never have bound, so the failure mode was UnboundLocalError rather than whatever was intended.

GraphBase.__init__ — the one with teeth

s is assigned only inside the edge loop, but the connectivity check after it reads s unconditionally. With no edges, nothing binds it.

Today that is saved by short-circuiting: the check is v.component is None or (v.component != s), and with no edges every vertex has component is None, so s is never evaluated. That stops holding the moment a caller passes vertices that already carry a component — then the first vertex reaches v.component != s and raises UnboundLocalError.

s = None preserves the behaviour exactly: a vertex whose component is not None still fails != s, so it is still reported unconnected. Pinned both ways, including the single-vertex short circuit.

tangents()

Repeated the final tangent by appending the loop variable after the loop, which requires the loop to have run — guaranteed only by an assert n >= 2, and asserts vanish under -O. Now appends T[-1], which also states the intent. Verified identical to the old function on every input tried (n = 2, 3, 5, 9).

mercator_layout

Bound cupy only inside its import guard and gated the GPU branch on a separate use_cupy flag. cp is not None says the same thing with one name instead of two.

GPU validation (this changes which branch runs, so it is not a CPU-only claim): on a GB10 with cudf 26.02.01 / cupy 13.6.0, the cupy branch equals the numpy branch to 1e-9 across seven points including the equator, ±89° latitude and ±179.9° longitude.

x gpu: [-1362.76653 -823.831024 -1.422663 1683.25423 0. 2002.63764 -2002.63764]
x cpu: [-1362.76653 -823.831024 -1.422663 1683.25423 0. 2002.63764 -2002.63764]
MERCATOR_GPU_PARITY_OK

Its two cuDF-only lines carry the established # pragma: no cover - cuDF-only ... (validated on dgx), since no lane feeding the changed-line gate can execute them.

A smaller form that avoided those two lines entirely (cp = None then import cupy as cp) was tried and rejected: mypy reports no-redef.

Verification

bin/lint.sh (ruff, type-hygiene 4448, comment-density 1151) · pyright guard 255 · mypy 351 files · changed-line coverage 4/4 = 100% · 251 layout tests · the meta-guards, since this adds a test file · both mutable tests confirmed discriminating by mutating the code under them.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ropeBMLJUuy6ViYwy15ud