Partial cache-backed prediction batches

Implementation specification for #304 and compatibility audit for #374.

The default remains strict: a missing peptide, context, kind or genotype raises CachedPredictorCoverageError. Existing raise_on_error only controls variant annotation and must not silently change prediction completeness.

An explicit TopiaryPredictor(cache_miss_handler=...) opts into retaining successful model/input pairs. The handler receives a structured record for every skipped pair, with input identity, model identity, prediction stage and original error. Split only failed batches until the failing input is isolated. Skip the entire sequence for that model, not just its missing window: this preserves protein flanks and never substitutes an incompatible prediction. Other models can still score that sequence. Whole-peptide, fragment, wildtype and self-nearest prediction paths use the same policy; unexpected errors still propagate.

A public predict_with_cache_miss_report function supplies this shared policy. The handler is the authoritative report across composed workflows; successful rows remain ordinary prediction rows. Emit an explicit warning when any inputs are skipped. Callers must retain the report alongside the partial table.

CLI --cache-miss-report PATH opts in, requires a real JSON file distinct from inputs and prediction outputs, and writes a versioned completeness report before writing the result table. Complete runs return 0; partial runs return 3, including runs with zero successful predictions. CSV stdout remains parseable. The report is written even if every cache lookup succeeds. No predictions are fabricated.

Tests must run strict and reporting modes on the same covered/uncovered inputs, assert successful sequences/models survive, keep repeated peptide contexts separate, preserve kind/genotype failure checks, exercise composed fragment/WT and CLI workflows, and prove unrelated exceptions and handler failures still raise. Fixture builders belong beside the tests.

Topiary 5.68.4 (#380) now delegates fixture resolution and generation to the published Osteosarc 0.2.3 fixture_assets and consumer_fixtures APIs. Osteosarc 0.1.4 lacks these APIs and is no longer a compatible development baseline. Topiary requires one osteosarc minor series at a time (see requirements.txt), accepting patch fixes within it. Verify shared fixtures offline, built-distribution metadata, normal dependency resolution and an editable installation. Shared source data and historical scientific expectations stay pinned independently of the package version.