[jayhesselberth]

Ports #11's narrow tail stripe to the half2 kernel: stripe2 is now a template on RR ∈ {16, 8, 4} and the last stripe takes the narrowest that covers the read's remaining rows (rows ≤ 4 → 4, ≤ 8 → 8), so a 130 nt read computes 132 rows instead of 144. The int kernel, MAX_REF_LEN_H2, fits_f16 and the launch bounds are untouched.

Measured (examples/gpu_score_probe.rs, one A30, M1 reads vs the 164-reference panel, batch 50,000, auto, arms interleaved, two reps each within 0.3%; fixed-length sets are M1 reads truncated to that length, 100,000 each):

reads	tail rows (stripe)	local main → tail	×	semi-global main → tail	×
100 nt	4 (4)	5.02e11 → 5.57e11	1.11	5.48e11 → 6.21e11	1.13
112 nt	16 (16)	5.61e11 → 5.62e11	1.00	6.12e11 → 6.30e11	1.03
120 nt	8 (8)	5.27e11 → 5.59e11	1.06	5.75e11 → 6.27e11	1.09
128 nt	16 (16)	5.61e11 → 5.63e11	1.00	6.13e11 → 6.31e11	1.03
130 nt	2 (4)	5.07e11 → 5.51e11	1.09	5.55e11 → 6.16e11	1.11
136 nt	8 (8)	5.30e11 → 5.60e11	1.06	5.80e11 → 6.28e11	1.08
144 nt	16 (16)	5.60e11 → 5.63e11	1.00	6.12e11 → 6.33e11	1.03
160 nt	16 (16)	5.60e11 → 5.62e11	1.00	6.13e11 → 6.32e11	1.03
M1, 100–200 nt	—	5.28e11 → 5.49e11	1.04	5.77e11 → 6.15e11	1.07
M1, all	—	5.25e11 → 5.48e11 (2.61 → 2.50 s)	1.04	5.74e11 → 6.13e11 (2.39 → 2.24 s)	1.07

Registers: __launch_bounds__(256, 3) unchanged, half2 kernel still 80 registers; local memory per thread 56 → 72 B (the int kernel holds 120 B at 64). New GpuScorer::{local_bytes, local_bytes_h2} report it and the probe prints it.

Parity: gpu_parity 5/5 on both kernels, both modes, the five schemes (including the positive-mismatch and free-gap ones only half2 masking survives), every read length 1–40 around the stripe; with the rows ≤ 8 arm deliberately sent to the 4-row stripe, 3 of its 5 tests fail (e.g. a 600 nt read scores 247 against the oracle's 251), so the test sees a wrong tail. CLI GPU tests 16/16 including device_gpu_output_is_byte_identical and pair_gpu_matches_cpu (A30); fmt, clippy with and without --features gpu, workspace tests and doctests on rna.

Docs: cuda/mod.rs, cuda/kernel.rs, benchmarks/README.md (new section with the table and reproduce lines), CHANGELOG.md (Unreleased), crates/eschalign/CLAUDE.md.

Closes #13