Support externally generated variant annotations

Background

Varcode currently computes all variant effects internally via predict_variant_effects(), which walks PyEnsembl transcripts and predicts coding consequences from first principles. This works well for point variants where the genomic-to-protein mapping is deterministic, but breaks down for structural variants and complex events where:

• The effect on the protein product may not be inferrable from the genomic rearrangement alone
• External tools (e.g., Exacto, long-read RNA assemblers) have already determined the transcript structure and translated the resulting protein
• Annotation requires RNA-level evidence that varcode doesn't have access to

Scope

Allow varcode to accept pre-computed annotations from external tools, rather than requiring all effects to be internally predicted.

Design considerations

1. Annotation source tracking: effects should record how they were determined — internally predicted vs. externally provided, and from which tool/pipeline.
2. Coexistence with prediction: for the same variant, a user might want both the internally predicted effect(s) and the externally provided one(s) (e.g., to compare DNA-only predictions against RNA-observed outcomes).
3. Structured import API: provide a way to attach an EffectCollection (or individual effects) to a variant from external data, with validation that the effect is compatible with the variant type.
4. Confidence/evidence metadata: external annotations may come with supporting evidence (read counts, transcript models, confidence scores) that should be preservable.
5. Sets of effects, not single annotations: a single variant (especially an SV) can produce multiple distinct protein products via alternative splicing. External annotations should be importable as a set of observed or possible effects for one variant, not forced into a single "the effect" slot. This is especially important when importing Exacto results, where one DNA variant may link to multiple transcript models, each with its own primary structure.

Relevant context

• Exacto produces annotated variant calls with gene/transcript/exon annotations (position_1_annotation, position_2_annotation columns) and per-amino-acid primary structure tables that trace each residue back to its source variant. Critically, Exacto's integration table maps one dna_variant_call_id to potentially many transcript_model_id entries, reflecting the multiple spliced isoforms a single DNA event can produce.
• This is a prerequisite for meaningfully importing Exacto structural variant results, since varcode cannot internally predict the effect of most SVs.

Start with predict_variant_effects() and the existing EffectCollection concepts, then review the Exacto fields position_1_annotation, position_2_annotation, and its transcript_model_id mappings. Done means externally supplied sets of effects can be attached to a variant, retain source and evidence metadata, coexist with predictions, and validate compatibility with the variant type.