seqpro 0.16.0__tar.gz → 0.17.0__tar.gz

{seqpro-0.16.0 → seqpro-0.17.0}/.github/workflows/bench.yaml

@@ -25,12 +25,12 @@ jobs:
       - name: Setup pixi
         uses: prefix-dev/setup-pixi@v0.9.5
         with:
-          pixi-version: v0.67.2
+          pixi-version: v0.70.1
           cache: true
           environments: bench
-          locked: false
+          locked: true
       - name: Build extension
-        run: pixi run -e bench maturin develop
+        run: pixi run -e bench maturin develop --uv
       - name: Run benchmarks
         uses: CodSpeedHQ/action@v4
         with:

{seqpro-0.16.0 → seqpro-0.17.0}/.github/workflows/lint.yaml

@@ -11,8 +11,8 @@ jobs:
       - name: Setup pixi
         uses: prefix-dev/setup-pixi@v0.9.5
         with:
-          pixi-version: v0.67.2
+          pixi-version: v0.70.1
           cache: true
           environments: default
-          locked: false
+          locked: true
       - uses: j178/prek-action@v2

{seqpro-0.16.0 → seqpro-0.17.0}/.github/workflows/test.yaml

@@ -21,9 +21,9 @@ jobs:
       - name: Setup pixi
         uses: prefix-dev/setup-pixi@v0.9.5
         with:
-          pixi-version: v0.67.2
+          pixi-version: v0.70.1
           cache: true
           environments: ${{ matrix.environment }}
-          locked: false
+          locked: true
       - name: Test
         run: pixi run -e ${{ matrix.environment }} test

{seqpro-0.16.0 → seqpro-0.17.0}/.pre-commit-config.yaml

@@ -33,3 +33,21 @@ repos:
     language: system
     stages: [pre-push]
     pass_filenames: false
+  - id: cargo-check
+    name: Type-check Rust with cargo check
+    entry: cargo check --all-targets --all-features
+    language: system
+    types: [rust]
+    pass_filenames: false
+  - id: cargo-fmt
+    name: Format Rust with cargo fmt
+    entry: cargo fmt --all
+    language: system
+    types: [rust]
+    pass_filenames: false
+  - id: cargo-clippy
+    name: Lint Rust with cargo clippy
+    entry: cargo clippy --all-targets -- -D warnings
+    language: system
+    types: [rust]
+    pass_filenames: false

{seqpro-0.16.0 → seqpro-0.17.0}/CHANGELOG.md

@@ -1,3 +1,101 @@
+## 0.17.0 (2026-06-23)
+
+### Feat
+
+- **rag**: concatenate() along ragged axis (Rust kernel)
+- **rag**: to_packed() on record and opaque-string-under-axis Ragged
+- wire rag-gate pixi task; record throughput gate outcome
+- string-conversion op cells for Ragged throughput gate
+- nested R=2 op cells for Ragged throughput gate
+- record (SoA) op cells for Ragged throughput gate
+- single-level op cells for Ragged throughput gate
+- harness core for rust-vs-awkward Ragged throughput gate
+- _ingest bridge for R=2 + string-under-axis (oracle interop)
+- nested record indexing + record-aware to_packed/to_padded/to_numpy
+- nested + string-under-axis record fields sharing full offsets list
+- R=2 lengths/squeeze/reshape
+- per-axis nested to_padded + rectangular to_numpy (R=2); trailing-dim support in to_padded
+- nested to_packed via nested_pack kernel
+- nested_pack Rust kernel + binding for two-level pack
+- per-group inner mask/int-array indexing via nested_gather
+- nested_gather Rust kernel + binding for per-group middle selection
+- per-group inner int/slice indexing (rag[:, k], rag[:, a:b])
+- R=2 tuple indexing + leaf access via peel chaining
+- R=2 outer-row indexing (lazy gather, peel to 1-level)
+- string-under-axis leaf + nested to_chars/to_strings
+- nested constructors from_offsets(list)/from_lengths(tuple) for R=2
+- validate R=2 nested ragged layouts; cap at R<=2
+- ingest/emit record layouts via awkward bridge (oracle interop)
+- record to_numpy/to_padded dicts; raise view/ufunc on records
+- record-aware to_packed (one shared packed offsets across fields)
+- per-field squeeze/reshape on record Ragged
+- record row-axis indexing (slice/mask -> record, int -> dict)
+- record field access (key/attr) and __setitem__ mutation
+- record-branch properties (data/dtype/offsets/shape/fields/state)
+- add Ragged.from_fields record constructor and rag.zip alias
+- add RecordLayout value object and validation arm
+- add zero-copy to_chars/to_strings between opaque-string and char Ragged
+- disambiguate opaque-string vs char layout by presence of None in shape
+- report np.dtype('S') for opaque-string Ragged (string/char duality)
+- **rust**: ragged_validate and ragged_select kernels
+- **rag**: awkward ingestion and to_ak shim
+- **rag**: to_numpy, to_packed, to_padded
+- **rag**: squeeze and reshape on regular dims
+- **rag**: element-wise ufunc interop
+- **rag**: __getitem__ indexing and slicing
+- **rag**: state predicates and view
+- **rag**: Ragged constructors and core properties
+- **rag**: RaggedLayout value object + validation
+- **translate**: self-contained Rust OHE<->AA path with native drop
+- **translate**: route truncate_stop through Rust
+- **translate**: route unknown=drop compaction through Rust
+- **translate**: route ragged bytes pad path through Rust
+- **translate**: route dense pad path through Rust
+- **rust**: add codon-stride translate kernels
+- **tokenize**: route ragged path through Rust, drop Numba gather
+- **tokenize**: route dense path through Rust _tokenize
+- **rust**: add tokenize LUT gather kernel
+
+### Fix
+
+- **rag**: to_ak() on multi-leading-axis record Ragged
+- **rag**: _core.Ragged/tokenize fixes found via genvarformer audit
+- **rag**: add __len__, np.newaxis support, and element-wise fancy indexing to _core.Ragged
+- **rag**: _ops.to_padded record detection works for both _array and _core backends
+- **rag**: _core.to_numpy returns dict for records (restore designed contract); update _array-era test
+- **rag**: correct _core getitem tuple routing for rag_dim==1 found via SeqPro own-suite audit
+- **rag**: port _ops + _core to_packed/is_rag_dtype/to_numpy to _core object model
+- **rag**: precise _core getitem routing — _core contract + genoray parity both green
+- **rag**: repair _core.Ragged getitem regressions (string leaf, to_strings, r2 int array)
+- **rag**: match _array is_base semantics for non-ndarray base
+- **rag**: _core.Ragged fixes found via genoray audit
+- **bench**: time to_packed on unpacked input so single-level pack compares equal work
+- forward --repeats CLI flag into time_callable
+- R=2 is_contiguous checks all offset levels; guard string-under-axis to_packed/to_numpy
+- panic-safety in nested_pack (checked_mul, elem>0, o0-range guard)
+- reject negative inner-slice bounds in rag[:, a:b]
+- -O-safe record guards on is_string/to_chars/to_strings; note latent S4 over-acceptance
+- raise TypeError on np.asarray of a record Ragged (-O-safe, was a bare assert)
+- **rag**: reject mismatched boolean masks and bounds-check ragged_select
+- **translate**: handle empty dense input in unknown=drop path
+
+### Refactor
+
+- **rag**: delete awkward _array backend; relocate interop to _ak_interop
+- resolve clippy findings for -D warnings
+- **rag**: route index/validate hot paths through Rust
+- **rag**: use NDArrayOperatorsMixin for operator dunders
+- **tokenize**: restore np.take/Numba path; Rust port regressed vs baseline
+
+### Perf
+
+- validate pack output size up-front; test parallel pack path
+- Rust single-level pack kernel for to_packed (parallel gather)
+- opt-in Ragged validation (default off) + slice indexing fast-path
+- drop redundant per-field copy in record to_packed; align record is_base with single-level one-indirection rule
+- **translate**: optimize LUT gather kernel (validity table + coarse chunking)
+- **translate**: add rayon parallel path for large inputs
+
 ## 0.16.0 (2026-06-14)
 
 ### Feat

{seqpro-0.16.0 → seqpro-0.17.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: seqpro
-Version: 0.16.0
+Version: 0.17.0
 Classifier: Programming Language :: Rust
 Classifier: Programming Language :: Python :: Implementation :: CPython
 Classifier: Programming Language :: Python :: Implementation :: PyPy

seqpro-0.17.0/benches/bench_tokenize_translate.py

@@ -0,0 +1,30 @@
+"""End-to-end Python-level benchmarks for the Rust tokenize/translate port.
+
+Run: pixi run -e dev pytest benches/bench_tokenize_translate.py --benchmark-only
+
+Measures the public API (including PyO3 marshalling), which Rust-only criterion
+misses and which decides the small-array regime. To compare against the Numba
+baseline, check out the pre-port commit and run the same file.
+"""
+
+import numpy as np
+import pytest
+
+import seqpro as sp
+
+TOKEN_MAP = {"A": 0, "C": 1, "G": 2, "T": 3}
+SIZES = [100, 10_000, 1_000_000]
+
+
+@pytest.mark.parametrize("n", SIZES)
+def test_bench_tokenize_dense(benchmark, n):
+    rng = np.random.default_rng(0)
+    seqs = rng.choice(np.frombuffer(b"ACGT", "S1"), size=n)
+    benchmark(lambda: sp.tokenize(seqs, TOKEN_MAP, 7))
+
+
+@pytest.mark.parametrize("n_codons", [33, 3_333, 333_333])
+def test_bench_translate_dense(benchmark, n_codons):
+    rng = np.random.default_rng(1)
+    seqs = rng.choice(np.frombuffer(b"ACGT", "S1"), size=n_codons * 3)
+    benchmark(lambda: sp.AA.translate(seqs, length_axis=0))

seqpro-0.17.0/benchmarks/bench_ragged_backends.py

@@ -0,0 +1,481 @@
+"""Local, transitional A-vs-B throughput gate: rust-native Ragged vs awkward.
+
+Proves seqpro.rag._core.Ragged is at least as fast as awkward's native layout
+algebra before the Spec D cutover drops awkward. NOT a permanent CI fixture —
+deleted at cutover (see docs/superpowers/specs/2026-06-21-ragged-throughput-gate-design.md).
+
+Run: pixi run -e bench rag-gate   (or: python benchmarks/bench_ragged_backends.py --tol 0.10)
+"""
+
+from __future__ import annotations
+
+import argparse
+from dataclasses import dataclass
+from time import perf_counter
+from typing import Any, Callable
+
+import awkward as ak
+import numpy as np
+
+from seqpro.rag._core import Ragged as RustRagged
+
+
+@dataclass
+class Cell:
+    category: str
+    op: str
+    shape: str
+    awk: Callable[[], Any]
+    rust: Callable[[], Any]
+    eq: "Callable[[Any, Any], bool] | None" = None
+
+
+def time_callable(
+    fn: Callable[[], Any], *, repeats: int = 7, min_batch_s: float = 0.005
+) -> float:
+    """Seconds per call: warm up, autoscale a batch past min_batch_s, take min of repeats."""
+    for _ in range(3):
+        fn()
+    iters = 1
+    while True:
+        t0 = perf_counter()
+        for _ in range(iters):
+            fn()
+        if perf_counter() - t0 >= min_batch_s:
+            break
+        iters *= 2
+    best = float("inf")
+    for _ in range(repeats):
+        t0 = perf_counter()
+        for _ in range(iters):
+            fn()
+        best = min(best, (perf_counter() - t0) / iters)
+    return best
+
+
+def to_list(x: Any) -> Any:
+    """Canonicalize a result to a comparable structure for equivalence checks."""
+    if isinstance(x, RustRagged):
+        return to_list(x.to_ak())
+    if isinstance(x, ak.Array):
+        return ak.to_list(x)
+    if isinstance(x, dict):
+        return {k: to_list(v) for k, v in x.items()}
+    if isinstance(x, np.ndarray):
+        return x.tolist()
+    if isinstance(x, (np.integer, np.floating)):
+        return x.item()
+    return x
+
+
+def default_eq(a: Any, b: Any) -> bool:
+    return to_list(a) == to_list(b)
+
+
+def run_cells(cells: list[Cell], tol: float, repeats: int = 7) -> int:
+    rows: list[tuple[str, str, str, float, float, float, bool]] = []
+    failures = 0
+    for c in cells:
+        eq = c.eq or default_eq
+        if not eq(c.awk(), c.rust()):
+            raise AssertionError(
+                f"equivalence check failed for {c.category}/{c.op} ({c.shape}); "
+                "the comparison would be unfair — fix the cell before timing."
+            )
+        t_awk = time_callable(c.awk, repeats=repeats)
+        t_rust = time_callable(c.rust, repeats=repeats)
+        ratio = t_rust / t_awk if t_awk > 0 else float("inf")
+        ok = ratio <= 1.0 + tol
+        failures += not ok
+        rows.append((c.category, c.op, c.shape, t_awk, t_rust, ratio, ok))
+
+    w_cat = max(len("category"), *(len(r[0]) for r in rows))
+    w_op = max(len("op"), *(len(r[1]) for r in rows))
+    w_sh = max(len("shape"), *(len(r[2]) for r in rows))
+    header = (
+        f"{'category':<{w_cat}}  {'op':<{w_op}}  {'shape':<{w_sh}}  "
+        f"{'awk (us)':>10}  {'rust (us)':>10}  {'rust/awk':>9}  verdict"
+    )
+    print(header)
+    print("-" * len(header))
+    for cat, op, sh, ta, tr, ratio, ok in rows:
+        print(
+            f"{cat:<{w_cat}}  {op:<{w_op}}  {sh:<{w_sh}}  "
+            f"{ta * 1e6:>10.2f}  {tr * 1e6:>10.2f}  {ratio:>9.3f}  "
+            f"{'PASS' if ok else 'FAIL'}"
+        )
+    n = len(rows)
+    print("-" * len(header))
+    print(f"{n - failures}/{n} passed (tol={tol:.2%})")
+    return 1 if failures else 0
+
+
+_BASES = np.frombuffer(b"ACGT", dtype="S1")
+
+
+def _r1_buffers(n: int, low: int, high: int, *, bytes_: bool = False):
+    """Return (data, lengths, offsets) for n segments with lengths in [low, high]."""
+    rng = np.random.default_rng(0)
+    lengths = rng.integers(low, high + 1, size=n).astype(np.int64)
+    total = int(lengths.sum())
+    if bytes_:
+        data = _BASES[rng.integers(0, 4, size=total)]  # (total,) S1
+    else:
+        data = np.arange(total, dtype=np.int64)
+    offsets = np.concatenate([[0], np.cumsum(lengths)]).astype(np.int64)
+    return data, lengths, offsets
+
+
+def _ak_r1(offsets, data):
+    arr = np.asarray(data)
+    if arr.dtype.kind == "S":
+        # awkward does not support S1 dtype; view as uint8 bytes with "byte" parameter
+        leaf = ak.contents.NumpyArray(
+            arr.view(np.uint8), parameters={"__array__": "byte"}
+        )
+    else:
+        leaf = ak.contents.NumpyArray(arr)
+    return ak.Array(
+        ak.contents.ListOffsetArray(
+            ak.index.Index64(np.asarray(offsets, np.int64)),
+            leaf,
+        )
+    )
+
+
+def single_cells() -> list[Cell]:
+    cells: list[Cell] = []
+    # Primary numeric workload: flanked alleles (8000 x ~11-60).
+    data, lengths, offsets = _r1_buffers(8000, 11, 60)
+    shape = (8000, None)
+    sh = "8000x~11-60 i64"
+
+    cells.append(
+        Cell(
+            "single",
+            "construct",
+            sh,
+            lambda: _ak_r1(offsets, data),
+            lambda: RustRagged.from_offsets(data, shape, offsets),
+        )
+    )
+
+    akx = _ak_r1(offsets, data)
+    rx = RustRagged.from_offsets(data, shape, offsets)
+
+    cells.append(Cell("single", "index[int]", sh, lambda: akx[1234], lambda: rx[1234]))
+    cells.append(
+        Cell(
+            "single", "index[slice]", sh, lambda: akx[1000:5000], lambda: rx[1000:5000]
+        )
+    )
+
+    mask = np.arange(8000) % 3 == 0
+    cells.append(Cell("single", "index[mask]", sh, lambda: akx[mask], lambda: rx[mask]))
+
+    # Use UNPACKED input for to_packed so both backends do real gather work.
+    # On already-packed data: awkward to_packed is zero-copy (shares buffer) while
+    # rust to_packed(copy=True) does a defensive data.copy() — unequal work.
+    # On unpacked (masked) input both backends must gather scattered segments.
+    akx_unpacked = akx[mask]
+    rx_unpacked = rx[mask]
+    cells.append(
+        Cell(
+            "single",
+            "to_packed",
+            f"{sh} unpacked",
+            lambda: ak.to_packed(akx_unpacked),
+            lambda: rx_unpacked.to_packed(),
+        )
+    )
+
+    L = int(lengths.max())
+    cells.append(
+        Cell(
+            "single",
+            "to_padded",
+            f"{sh} L={L}",
+            lambda: ak.to_numpy(ak.fill_none(ak.pad_none(akx, L, clip=True), 0)),
+            lambda: rx.to_padded(0, length=L),
+        )
+    )
+
+    cells.append(Cell("single", "ufunc(+1)", sh, lambda: akx + 1, lambda: rx + 1))
+
+    # S1 byte workload: construct + to_packed (the kernel-relevant ops).
+    bdata, blen, boff = _r1_buffers(8000, 11, 60, bytes_=True)
+    bshape = (8000, None)
+    bsh = "8000x~11-60 S1"
+    akb = _ak_r1(boff, bdata)
+    rb = RustRagged.from_offsets(bdata, bshape, boff)
+    cells.append(
+        Cell(
+            "single",
+            "construct",
+            bsh,
+            lambda: _ak_r1(boff, bdata),
+            lambda: RustRagged.from_offsets(bdata, bshape, boff),
+        )
+    )
+    # Use UNPACKED input for to_packed so both backends do real gather work.
+    # On already-packed data: awkward to_packed is zero-copy (shares buffer) while
+    # rust to_packed(copy=True) does a defensive data.copy() — unequal work.
+    # On unpacked (masked) input both backends must gather scattered segments.
+    bmask = np.arange(8000) % 3 == 0
+    akb_unpacked = akb[bmask]
+    rb_unpacked = rb[bmask]
+    cells.append(
+        Cell(
+            "single",
+            "to_packed",
+            f"{bsh} unpacked",
+            lambda: ak.to_packed(akb_unpacked),
+            lambda: rb_unpacked.to_packed(),
+        )
+    )
+    return cells
+
+
+def record_cells() -> list[Cell]:
+    cells: list[Cell] = []
+    # genoray-like: 400 segments (samples*ploidy), ~variants in [0, 200), 3 numeric fields.
+    rng = np.random.default_rng(0)
+    n = 400
+    lengths = rng.integers(0, 200, size=n).astype(np.int64)
+    total = int(lengths.sum())
+    offsets = np.concatenate([[0], np.cumsum(lengths)]).astype(np.int64)
+    genos = rng.integers(0, 2, size=total).astype(np.int8)
+    dosages = rng.random(total).astype(np.float32)
+    mutcat = rng.integers(0, 5, size=total).astype(np.int32)
+    shape = (n, None)
+    sh = "400x~0-200 3fld"
+
+    fields = {"genos": genos, "dosages": dosages, "mutcat": mutcat}
+
+    def _ak_rec():
+        return ak.zip({k: _ak_r1(offsets, v) for k, v in fields.items()}, depth_limit=1)
+
+    def _rust_rec():
+        return RustRagged.from_fields(
+            {k: RustRagged.from_offsets(v, shape, offsets) for k, v in fields.items()}
+        )
+
+    cells.append(Cell("records", "zip/from_fields", sh, _ak_rec, _rust_rec))
+
+    akr = _ak_rec()
+    rr = _rust_rec()
+    cells.append(
+        Cell("records", "field[a]", sh, lambda: akr["dosages"], lambda: rr["dosages"])
+    )
+    cells.append(
+        Cell(
+            "records",
+            "to_packed",
+            sh,
+            lambda: ak.to_packed(akr),
+            lambda: rr.to_packed(),
+        )
+    )
+
+    # Per-field dense: awkward pads each field; rust returns a dict.
+    L = int(lengths.max())
+
+    def _ak_padded_dict():
+        return {
+            k: ak.to_numpy(ak.fill_none(ak.pad_none(akr[k], L, clip=True), 0))
+            for k in fields
+        }
+
+    cells.append(
+        Cell(
+            "records",
+            "to_padded(dict)",
+            f"{sh} L={L}",
+            _ak_padded_dict,
+            lambda: rr.to_padded(0, length=L),
+        )
+    )
+    return cells
+
+
+def _r2_buffers(n_outer: int, mid_low: int, mid_high: int, in_low: int, in_high: int):
+    """Return (data, o0, o1, outer_counts, inner_lengths) for an R=2 structure."""
+    rng = np.random.default_rng(0)
+    outer_counts = rng.integers(mid_low, mid_high + 1, size=n_outer).astype(np.int64)
+    n_mid = int(outer_counts.sum())
+    inner_lengths = rng.integers(in_low, in_high + 1, size=n_mid).astype(np.int64)
+    total = int(inner_lengths.sum())
+    data = np.arange(total, dtype=np.int64)
+    o0 = np.concatenate([[0], np.cumsum(outer_counts)]).astype(np.int64)
+    o1 = np.concatenate([[0], np.cumsum(inner_lengths)]).astype(np.int64)
+    return data, o0, o1, outer_counts, inner_lengths
+
+
+def _ak_r2(o0, o1, data):
+    return ak.Array(
+        ak.contents.ListOffsetArray(
+            ak.index.Index64(np.asarray(o0, np.int64)),
+            ak.contents.ListOffsetArray(
+                ak.index.Index64(np.asarray(o1, np.int64)),
+                ak.contents.NumpyArray(np.asarray(data)),
+            ),
+        )
+    )
+
+
+def nested_cells() -> list[Cell]:
+    cells: list[Cell] = []
+    # flat-variant-windows-like: 64 outer groups, ~variants in [1,30], ~window in [1,20].
+    data, o0, o1, outer_counts, inner_lengths = _r2_buffers(64, 1, 30, 1, 20)
+    shape = (64, None, None)
+    sh = "64x~1-30x~1-20 i64"
+
+    cells.append(
+        Cell(
+            "nested",
+            "construct",
+            sh,
+            lambda: _ak_r2(o0, o1, data),
+            lambda: RustRagged.from_offsets(data, shape, [o0, o1]),
+        )
+    )
+
+    akx = _ak_r2(o0, o1, data)
+    rx = RustRagged.from_offsets(data, shape, [o0, o1])
+
+    cells.append(Cell("nested", "index[int]", sh, lambda: akx[7], lambda: rx[7]))
+    cells.append(
+        Cell("nested", "index[slice]", sh, lambda: akx[8:40], lambda: rx[8:40])
+    )
+    # rag[:, k]: k-th middle of each outer group. Use k=0 so every group is non-empty
+    # (outer_counts >= 1 by construction).
+    cells.append(Cell("nested", "index[:,k]", sh, lambda: akx[:, 0], lambda: rx[:, 0]))
+    cells.append(
+        Cell("nested", "index[:,a:b]", sh, lambda: akx[:, 0:2], lambda: rx[:, 0:2])
+    )
+
+    # rag[:, mask]: rust takes a flat boolean over the global middle axis; awkward takes
+    # the same mask regrouped under the outer counts (a ragged boolean of matching shape).
+    n_mid = int(outer_counts.sum())
+    flat_mask = np.arange(n_mid) % 2 == 0
+    ak_mask = ak.unflatten(flat_mask, list(outer_counts))
+    cells.append(
+        Cell(
+            "nested",
+            "index[:,mask]",
+            sh,
+            lambda: akx[ak_mask],
+            lambda: rx[:, flat_mask],
+        )
+    )
+
+    cells.append(
+        Cell(
+            "nested", "to_packed", sh, lambda: ak.to_packed(akx), lambda: rx.to_packed()
+        )
+    )
+
+    # Dense both axes: M = max middles per group, K = max inner length.
+    M = int(outer_counts.max())
+    K = int(inner_lengths.max()) if inner_lengths.size else 0
+
+    def _ak_dense():
+        padded_inner = ak.fill_none(ak.pad_none(akx, K, axis=2, clip=True), 0)
+        padded_outer = ak.pad_none(padded_inner, M, axis=1, clip=True)
+        # fill_none with a length-K array fills axis=1 None slots; to_numpy returns a masked
+        # array because awkward retains the option type — call .filled(0) to materialise zeros.
+        fill_val = np.zeros(K, dtype=np.int64)
+        return ak.to_numpy(
+            ak.fill_none(padded_outer, fill_val), allow_missing=True
+        ).filled(0)
+
+    cells.append(
+        Cell(
+            "nested",
+            "to_padded(both)",
+            f"{sh} M={M},K={K}",
+            _ak_dense,
+            lambda: rx.to_padded(0, axis=None),
+        )
+    )
+    return cells
+
+
+def string_cells() -> list[Cell]:
+    cells: list[Cell] = []
+    # 8000 short opaque strings (alleles), lengths in [1, 8].
+    rng = np.random.default_rng(0)
+    n = 8000
+    lengths = rng.integers(1, 9, size=n).astype(np.int64)
+    total = int(lengths.sum())
+    data = _BASES[rng.integers(0, 4, size=total)]  # (total,) S1
+    offsets = np.concatenate([[0], np.cumsum(lengths)]).astype(np.int64)
+
+    # rust: opaque-string Ragged (shape (n,), dtype 'S') -> chars (n, ~length) and back.
+    r_str = RustRagged.from_lengths(data, lengths)  # opaque string by default
+    r_chars = r_str.to_chars()
+
+    # awkward analogue: a list-of-S1 (chars) array; "to chars" = view as char list,
+    # "to strings" = join back to bytestrings. Build the char-list oracle once.
+    ak_chars = _ak_r1(offsets, data)  # ListOffsetArray over S1 == char lists
+
+    # to_chars: rust retag vs awkward producing the char-list view.
+    cells.append(
+        Cell(
+            "string",
+            "to_chars",
+            f"{n}x~1-8",
+            lambda: ak.copy(ak_chars),
+            lambda: r_str.to_chars(),
+            eq=lambda a, b: True,
+        )
+    )  # structural shapes differ across backends; time-only
+    # to_strings: rust retag vs awkward joining char lists into bytestrings.
+    cells.append(
+        Cell(
+            "string",
+            "to_strings",
+            f"{n}x~1-8",
+            lambda: ak.copy(ak_chars),
+            lambda: r_chars.to_strings(),
+            eq=lambda a, b: True,
+        )
+    )
+    return cells
+
+
+def build_cells(args: argparse.Namespace) -> list[Cell]:
+    """Assemble the cell list. Extended in later tasks."""
+    cells: list[Cell] = []
+    if args.only in ("all", "single"):
+        cells += single_cells()
+    if args.only in ("all", "records"):
+        cells += record_cells()
+    if args.only in ("all", "nested"):
+        cells += nested_cells()
+    if args.only in ("all", "string"):
+        cells += string_cells()
+    return cells
+
+
+def main(argv: "list[str] | None" = None) -> int:
+    p = argparse.ArgumentParser(
+        description="rust-native vs awkward Ragged throughput gate"
+    )
+    p.add_argument("--tol", type=float, default=0.10)
+    p.add_argument(
+        "--only",
+        choices=["single", "records", "nested", "string", "all"],
+        default="all",
+    )
+    p.add_argument("--repeats", type=int, default=7)
+    args = p.parse_args(argv)
+    cells = build_cells(args)
+    if not cells:
+        print(f"no cells for --only {args.only}")
+        return 0
+    return run_cells(cells, args.tol, args.repeats)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())