feat: configurable track insertion fill strategies

[d-laub]

Summary

• Adds 5 user-selectable strategies for filling track values at insertion sites: Repeat5p (default, byte-identical to prior behavior), Repeat5pNormalized, Constant, FlankSample, and Interpolate (linear/quadratic/cubic Lagrange).
• New Dataset.with_insertion_fill(strategy_or_dict) lets users set a strategy globally or per active track. Returns a new lazy view; original dataset unchanged.
• Strategy dispatch happens inside the numba kernel shift_and_realign_track_sparse. FlankSample uses an inline xorshift64 hash of (base_seed, query, hap, out_idx) — parallel-safe, no np.random globals. deterministic=True derives a reproducible base_seed from the idx array.

Design + Plan

• Spec: docs/superpowers/specs/2026-05-11-track-insertion-options-design.md
• Plan: docs/superpowers/plans/2026-05-11-track-insertion-options.md

Test Plan

• Kernel-level unit test for every strategy (tests/dataset/test_insertion_fill.py)
• Default-path regression guard — Repeat5p produces byte-identical output to pre-change behavior
• FlankSample determinism across calls + sensitivity to (query, hap) distinguishing seeds
• Edge clamping when flank pool extends beyond the track
• Per-track dict form falls back to Repeat5p for unspecified tracks
• with_tracks prunes insertion_fill to active tracks
• End-to-end through Dataset[...] with the dummy dataset
• Rejection when reconstructor isn't HapsTracks
• Full tests/dataset/ suite — 381 passed, 5 skipped, 1 xfailed, no regressions

Follow-ups (non-blocking, called out in final review)

• The deterministic base_seed uses bitwise_xor.reduce(idx), which is permutation-invariant. The per-position (query, hap, pos) mixing in the kernel mitigates this for fills, but a position-sensitive hash would be stronger.
• The _REPEAT_5P arm inside _apply_insertion_fill is unreachable from the outer kernel (which short-circuits) and is commented as such.

🤖 Generated with Claude Code