Experimental transcript_model classifies a cross-boundary coding DUP as FivePrimeUTR

Reproduction on the installed PyPI 9.2.2 wheel

from pyensembl import cached_release
from varcode import StructuralVariant
g = cached_release(81)
t = g.transcript_by_id("ENST00000003084")
v = StructuralVariant("7", t.start - 100, "DUP", end=t.end - 50, genome=g)
e = v.effect_on_transcript(t, annotator="transcript_model")
print(type(e).__name__)
print([type(c.effect).__name__ for c in e.candidates])
print(e.modifies_protein_sequence)

Actual: FivePrimeUTR, ['FivePrimeUTR'], False.

The event is chr7:117479863–117668615, crosses the transcript boundary and
includes the CFTR coding sequence. This does not prove a changed protein:
an intact copy can remain. But neither does it establish that the event is
solely a 5-prime UTR edit, or resolve the transcript created at its junction.

GenomicLayout.apply_structural_variant clamps the affected interval to its
available reference segments before applying DUP/INV. The transcript model
constructs a finite layout (50-base flanks), so the distal junction can be
outside that layout. This is a separate path from the default structural
builder corrected in #448 / #405. Current #448 code reproduces this same
experimental result; this is not a regression from that fix.

Focused fix boundary

Reject or explicitly leave unresolved nonlocal DUP/INV edits whose junctions
cannot be represented by the available layout, rather than treating a clipped
body as the entire event. Preserve explicit assembled alleles and fully
represented local events. Add both-strand, left/right/enclosing and isoform
coverage regressions. Do not replace this with a blanket protein-modifying flag.

Related broader classification/filtering design: #420, #418, #423.

Start at the transcript_model path and GenomicLayout.apply_structural_variant, then compare its finite-layout handling with the default structural builder discussed in #448 and #405. Trace DUP/INV cases where a junction falls outside the available layout, including both strands, left/right/enclosing events, and multiple isoforms. Done means nonlocal edits are rejected or left unresolved while explicit assembled alleles and fully represented local events retain their behavior.