Generalize phasing API; remove Isovar-named identifiers

[iskandr]

Summary

Closes #378. Filed companion isovar tickets: openvax/isovar#184.

Varcode's phasing API named one upstream tool (Isovar) on every public identifier — IsovarPhaseResolver, IsovarAssemblyProvider — even though varcode imports nothing isovar-shaped and isovar already declares varcode>=4.15. This made the dependency direction read as circular, forced non-Isovar sources to lie about their identity, and bundled two unrelated concerns (read-phasing + mutant-transcript construction) into a single Protocol.

This PR makes varcode's phasing surface fully generic. Implementations of the Protocols live in the packages that produce the underlying evidence (Isovar, long-read tools, custom assemblers, test stubs).

Renames

Before	After
IsovarAssemblyProvider (single Protocol)	Removed. Split into two narrower Protocols.
has_contig(variant, transcript)	has_evidence(variant) on ReadPhasingSource
variants_in_contig(variant, transcript)	partners_in_cis(variant) on ReadPhasingSource
mutant_transcript(variant, transcript)	Unchanged signature, now on MutantTranscriptSource
IsovarPhaseResolver	ReadPhaseResolver
resolver.source == "isovar"	resolver.source == "read_phasing"

ReadPhaseResolver requires a ReadPhasingSource. The mutant_transcript(...) method on the resolver is optional — it returns None silently when the wrapped source doesn't also satisfy MutantTranscriptSource, instead of raising. So a phasing-only source (e.g. the new isovar.IsovarReadPhasing from openvax/isovar#183) plugs in cleanly without needing to fabricate transcript reconstruction.

Why these names

• ReadPhasingSource — generic across data sources (Isovar, long-read tools, hand-rolled). Doesn't pretend any one upstream is special.
• has_evidence / partners_in_cis — neutral. The old has_contig was Isovar-vocabulary; long-read tools don't have "contigs," they have fragments. The phasing question is just "do you have evidence for this variant?" and "who's in cis with it?".
• MutantTranscriptSource — separate concern. A consumer that only needs phasing shouldn't be forced to also build transcripts.

Dependency direction stays one-way

• varcode/: zero import isovar. Zero Isovar-named identifiers on the public surface. Confirmed via grep.
• isovar/: continues to declare varcode>=4.15.0 and ship implementations of the new generic Protocols (Add IsovarReadPhasing adapter for RNA-read variant phasing isovar#183 already does this; Split VarcodeAdapter into two narrower providers; drop Varcode-named API isovar#184 covers the heavier VarcodeAdapter rework).

Files touched

• varcode/phasing.py — full rewrite of Protocol definitions + ReadPhaseResolver. VCFPhaseResolver and helpers untouched.
• varcode/__init__.py — updated re-exports + __all__.
• varcode/variant.py, varcode/variant_collection.py — phase_resolver= docstrings updated.
• tests/test_isovar_phase_resolver.py → tests/test_read_phase_resolver.py (git rename + body update). 18 tests, all on the new Protocol shape.
• tests/test_haplotype_effect.py — stub provider renamed to _StubReadPhasingSource matching the new shape; 4 tests on the resolver-driven path.
• tests/test_vcf_phase_resolver.py — interface-symmetry test now references ReadPhaseResolver.
• docs/germline.md — recipe and resolver list updated.
• docs/api.md — adds entries for ReadPhasingSource and MutantTranscriptSource; renames the resolver entry.
• CHANGELOG.md — Breaking entry.
• varcode/version.py — 4.23.0 → 4.24.0.

Test plan

• pytest tests/test_read_phase_resolver.py — 18 cases (Protocol conformance, phasing semantics, optional MutantTranscript channel, effect-attachment end-to-end). All pass.
• pytest tests/test_haplotype_effect.py — 7 cases (joint cDNA building, HaplotypeEffect emission via the new resolver). All pass.
• pytest tests/test_vcf_phase_resolver.py — 25 cases, unchanged behavior. All pass.
• pytest tests/test_germline_effects.py — 34 cases, unchanged behavior. All pass.
• pytest tests/ — 1188 passed, 2 pre-existing skips.

Out of scope

• The isovar-side rework of VarcodeAdapter — tracked in Split VarcodeAdapter into two narrower providers; drop Varcode-named API isovar#184. Until that lands, downstream users of VarcodeAdapter see a runtime error when wiring it into ReadPhaseResolver (the methods no longer match the Protocol). The new isovar adapter PR is small.
• Renaming IsovarPhaseResolver is the only public-surface rename in this PR; lower-level helper names are unchanged.

References

• Make phasing API generic: drop Isovar-specific names from public surface #378 — this PR's umbrella issue.
• Rename PhaseAmbiguousEffect -> PhaseCandidateSet #376 — recent PhaseAmbiguousEffect → PhaseCandidateSet rename (same hard-rename precedent).
• Germline-aware effect prediction (umbrella) #268 — germline-aware annotation umbrella.
• Add IsovarReadPhasing adapter for RNA-read variant phasing isovar#183 — IsovarReadPhasing already matches ReadPhasingSource.
• Split VarcodeAdapter into two narrower providers; drop Varcode-named API isovar#184 — companion isovar work for the heavy VarcodeAdapter rework.