0.2.0 — 2026-06-09

Release Notes

Real-prediction parity follow-up to 0.1.0. The headline work is six
new SOTA-harness cells that drive every kernel — bbox, segm, boundary,
keypoints, panoptic PQ, semantic mIoU, LVIS, and calibration — through
a frozen real-model prediction cache so the parity surface no longer
relies solely on synthetic fixtures. Two strict-mode behavioural fixes
ride along: the TIDE Missed-bin rewrite (previously a no-op under
parity_mode="strict") and the accumulator's n_d==0 precision/scores
write. The minor bump signals those output-value changes; the kernel
surface is otherwise unchanged from 0.1.0.

Added

• Real-prediction SOTA harness — six new cells. Each cell drives a
  pinned upstream checkpoint through the existing _harness_common
  scaffolding (full-SHA cache key, _ensure_pinned_revision preflight,
  torch.set_num_threads(1), int64 target_sizes, loud-fail on
  unmapped class names) and asserts vernier-vs-oracle parity on real
  output distributions:
  • DETR-R50 (#265) — instance bbox / segm against the
    facebook/detr-resnet-50 checkpoint on COCO val2017. Aligned tier
    loosens dtScores to rtol = 2 * eps to absorb the documented
    serde_json vs Python strtod 1-ULP score-parser drift; all
    integer-reduction surfaces (precision, recall, counts, 12-stat AP/AR
    summary) stay bit-equal.
  • Mask2Former panoptic + ADE-semantic (#266) — panoptic PQ
    against facebook/mask2former-swin-large-coco-panoptic on COCO
    panoptic val2017; semantic mIoU against the ADE checkpoint on
    ADE20K val. Both bit-equal to their oracles on integer-reduction
    surfaces.
  • DETR-R50 calibration (#267) — reuses the #265 prediction
    cache to validate ADR-0018 ECE / MCE / reliability against the
    NumPy oracle at full distribution scale.
  • rfdetr-segnano boundary (#269) — boundary IoU against
    bowenc0221's boundary_iou_api over the rfdetr-segnano TIDE cache;
    no new inference (boundary IoU is a different metric over the same
    RLE masks).
  • LVIS detector (#270) — federated LVIS evaluation against the
    LVIS API. Reuses the TIDE cache pattern; gates the K=168/817
    full-val divergence currently tracked in open_followups.md.
  • ViTPose keypoints (#271) — keypoints OKS evaluation against
    the usyd-community/vitpose-base-coco checkpoint on COCO val2017.

Fixed

• TIDE Missed-bin strict-mode parity (#273) — the rewrite-layer
  Missed fix was setting ignore_flag = Some(true) on missed GTs and
  relying on effective_ignore to resolve under both parity modes.
  Quirk D1's strict disposition discards ignore_flag entirely and
  reads only is_crowd, so under parity_mode="strict" the rewrite
  was a no-op: the AP denominator stayed unchanged and the per-bin
  delta collapsed to exactly 0.0 (vs the ADR-0021 NumPy oracle's
  spec'd 0.119 on DETR-R50). Fixed by deleting missed GTs from the
  corrected dataset entirely — parity-mode-independent and
  AP-equivalent to ignoring on the oracle's semantics. Validated to
  within 1 ULP against the oracle on COCO val2017 + DETR-R50
  (~150k detections, 8 ULP gate). Closes the ADR-0022 follow-up on
  t_b = 0.1 for set-prediction transformer detectors.
• n_d == 0 precision/scores write (#272) — the accumulator path
  for classes with zero detections now writes 0.0 (not -1) into
  the precision and scores tensors. Downstream consumers comparing
  raw tensor values across releases will see this change; the public
  AP / AR summary statistics are unaffected (they already skipped
  -1 sentinel entries).

Install vernier-cli 0.2.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/NoeFontana/vernier/releases/download/v0.2.0/vernier-cli-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/NoeFontana/vernier/releases/download/v0.2.0/vernier-cli-installer.ps1 | iex"

Download vernier-cli 0.2.0

File	Platform	Checksum
vernier-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
vernier-cli-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
vernier-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
vernier-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

0.1.0 — 2026-05-19

Release Notes

First release out of the 0.0.x line. Mostly a performance + parallelism
follow-up to 0.0.4 — no new evaluation paradigms, every shipped kernel
keeps strict bit-equal parity with its oracle. The cross-paradigm
benchmark page is refreshed against the post-0.0.4 SHA on the same
machine fingerprint as the 0.0.4 snapshot (37652a58e939).

Added

• num_threads parallelism (ADR-0047)
  (#251, #253, #254, #256) — opt-in num_threads: int | None = None
  on every public evaluate surface across all four paradigms: instance
  (bbox / segm / boundary / keypoints), semantic, panoptic, and LVIS,
  on batch + streaming + background entry points (Evaluator.evaluate,
  Evaluator.background, submit / submit_png). The sequential
  path (num_threads=None or 1) is byte-for-byte unchanged from
  0.0.4; no rayon symbol is entered. parity_threads parity tests
  assert bit-equal results across num_threads ∈ {None, 1, 2, 4, 8}
  on every paradigm. CLI gains vernier eval --threads N.
• bench-timings Cargo feature (#256) — atomic (par_iter, serial_post) split + build_*_anns call counter on
  evaluate_with_parallel, attributed via the new BenchCounterSet
  shared helper (#258). Off by default and stripped from the shipped
  wheel; powers the bbox-scaling attribution at
  docs/engineering/benchmarking/2026-05-bbox-cdf.md.
• mimalloc-global Cargo feature on vernier-ffi (#256) —
  allocator A/B knob, off by default; lets users opt into mimalloc
  for hot-allocation workloads without it being a default cost.
• Semantic divan microbench (#261) —
  crates/vernier-semantic/benches/accumulate_confusion.rs
  exercises three input distributions (realistic_perfect,
  realistic_jittered, uniform_random) at the val2017
  panoptic-semantic geometry; prereq for the chunked-u8 kernel work.

Changed

• bbox AP perf (#256, #258, #259) — KernelScratch per-worker
  annotation pool + direct-write parallel runner (replaces the
  per-image Vec<CellOutput> intermediate with par_chunks_mut);
  in-place image-major → canonical transpose via cycle-following
  (eliminates a 26 MB intermediate buffer pair on val2017); the
  eval_imgs + eval_imgs_meta transposes fuse into a single
  cycle walk (halves index arithmetic, drops one of two 1.6 MB
  visited-bitset allocations). Net val2017 nt=4: par_iter region
  42 → 32 ms, serial_post 45 → 19 ms, peak working-set
  −24 MB. The remaining Amdahl floor on --num-threads for bbox is
  the ~200 ms single-threaded dataset_build (HashMap validation in
  CocoDataset::from_parts), attributed via bench-timings.
• Panoptic PQ perf (#260) — sparse-remap adjacent-pixel cache on
  build_dense_intersections and build_dense_boundary_intersections.
  COCO panoptic always hits the sparse branch (RGB-packed ids exceed
  the 1 M dense cap) and panoptic segments are spatially contiguous,
  so consecutive (g, d) pairs are usually identical; a 4-state
  (last_g, last_d, last_gi, last_di) cache skips the FxHashMap
  lookup on adjacent-pixel matches. Dense branch is deliberately
  uncached (Vec::get is cheap enough that the miss overhead
  regresses synthetic by ~70%). SSSE3 RGB→u32 pack on the panoptic
  PNG decode path. New coco_like_rgb microbench arm exercises the
  sparse-RGB path that the existing coco_like arms missed
  (their ids 1..=50 took the dense path).
• Semantic mIoU perf (#261) — decode buffer pool + chunked u8
  kernel on accumulate_confusion for the T = u8 PNG fused-decode
  path that drives Semantic — mIoU (val2017). The pool reuses the
  per-image decode Vec<u8> across submissions; the chunked kernel
  keeps the strict-mode u64-additive fold but processes pixels in
  cache-line-sized batches.
• Background-evaluator threading wired (#253, #254) —
  BackgroundConfig.num_threads is no longer hardcoded None on the
  panoptic and semantic FFI ctors; BackgroundCapable gains a
  default-method apply_update_parallel that the panoptic and
  semantic streaming impls override. Panoptic submit_png defers
  PNG decode into the worker pool (PyBackedBytes zero-copy) so
  libpng decode parallelises across submissions; the single-threaded
  path keeps inline decode and is byte-for-byte unchanged.
• vernier-pixel-pack folded into vernier-panoptic — the
  SSSE3 RGB→u32 pack primitive added in #260 lived briefly as a
  standalone workspace crate. With a single consumer
  (vernier-panoptic::decode) and 172 LOC, it sat below the
  leaf-crate threshold and the audited-unsafe carveout fits cleanly
  inside the host crate (#![deny(unsafe_code)] at root, module-local
  #[allow(unsafe_code)] on the SSSE3 pshufb fn). Folding it
  back keeps the published crate set at the six 0.0.4 crates and
  avoids the registry-reservations + Trusted-Publisher loop in the
  release runbook for a non-reusable internal SIMD primitive.
• Bench harness --num-threads (#251, #252) — bench run --num-threads "1,2,4,8" override overrides the workload's pinned
  num_threads tuple; panoptic + semantic spawn helpers now forward
  the flag (previously dropped, so every panoptic / semantic cell
  ran with args.num_threads = None regardless of what the CLI
  swept).
• Bench page refreshed against 3a509df6c525 on the same
  37652a58e939 fingerprint as the 0.0.4 snapshot, so the speedup
  deltas are not confounded by host change. Per-cell movements
  (vernier median, 0.0.4 → HEAD):
  • panoptic PQ: 12.59 s → 10.53 s (−16.4%; speedup
    2.73× → 3.30× vs panopticapi). IQR also narrows from 21.22%
    to 9.78% (still over the 5% gate — PNG decode is chronically
    noisy on this host).
  • semantic mIoU val2017: 5.00 s → 2.82 s (−43.6%;
    speedup 4.12× → 7.40× vs mmsegmentation).
  • instance bbox / segm / boundary / keypoints / synth-semantic /
    LVIS move within VPS noise of their 0.0.4 numbers; speedups
    widen by 0.1×–0.5× as baselines drift slightly slower on this
    run.

Fixed

• bench run --impl all on non-instance paradigms — impls_for_iou
  raised KeyError for the paradigm-specific impls
  (vernier_panoptic, panopticapi, mmsegmentation,
  vernier_lvis, lvis-api) that #252 widened ALL_IMPLS to
  include. Falls back to an empty IoU set for impls that aren't
  registered for the instance paradigm.

Install vernier-cli 0.1.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/NoeFontana/vernier/releases/download/v0.1.0/vernier-cli-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/NoeFontana/vernier/releases/download/v0.1.0/vernier-cli-installer.ps1 | iex"

Download vernier-cli 0.1.0

File	Platform	Checksum
vernier-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
vernier-cli-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
vernier-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
vernier-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

0.0.4 — 2026-05-16

Release Notes

Robustness follow-up to 0.0.3. No new evaluation paradigms or kernel
changes — this release widens the typed-error surface, adds a fuzz
harness, and pins the platform-compat matrix in CI. The cross-paradigm
benchmark page is refreshed against the post-0.0.3 SHA on a single
fingerprint (no more dual-SHA LVIS caveat).

Added

• Typed Python error surface (#249) — four new PyValueError
  subclasses (InvalidAnnotationError, NonFiniteError,
  DimensionMismatchError, InvalidConfigError) with the public
  surface pinned by tests/python/test_error_matrix.py and documented
  at docs/reference/errors.md. Previously these all surfaced as bare
  ValueError; existing except ValueError: catches still match the
  new subclasses.
• Fuzz harness (#249) — tools/fuzz/ cargo-fuzz targets for the
  COCO / manifest / RLE / segmentation parsers (non-workspace crate so
  the nightly cargo-fuzz toolchain stays out of the publishable
  workspace). The vernier_core::fuzz_regressions integration test
  replays minimised crashes on every cargo nextest run; CI's
  slow.yml carries a 120 s/target smoke that builds once and exits.
• Platform-compat matrix (#249) — slow.yml adds a
  py3.10 × py3.13 × py3.14 ladder crossed with numpy / torch combos,
  exercising the BackgroundEvaluator tutorial end-to-end. Catches
  ABI / DLPack regressions that the single-version ci.yml matrix
  doesn't surface.
• bench-histogram Cargo feature (#249) — opt-in (G, D, wall_ns)
  per-call recorder on match_image, off by default and stripped from
  the shipped wheel. Powers the 10× val2017 scaling proof at
  docs/engineering/matching-scaling.md. Gated on
  vernier-core / vernier-ffi / vernier-mask; no production cost.
• Stress-matrix workloads (#249) — 6 named regimes
  (coco-baseline, detr-output, lvis-crowded, open-images-cats,
  satellite-4k, pathology-8k) plus per-axis sweeps in
  bench/workloads/stress_matrix.py; runner at
  bench/runners/stress_runner.py. Catalogue and expected behaviour
  per axis in docs/engineering/stress-matrix.md.
• Memory-under-training-load runner (#249) —
  bench/bench/runners/memory_bench.py (reuses
  bench.harness.rss.RSSSampler); methodology and reading guide at
  docs/engineering/memory-under-training.md.
• Colab smoke notebook (#249) — free-tier platform-check entry
  point; README badge links to it.

Changed

• Tutorial smoke now ingests via the DLPack array path —
  fake_model(image_ids) -> list[Detections] returning numpy arrays
  submitted batch-mode (matches torchvision's detection-API
  convention). Notebook cell-3 stays byte-identical to the .py body
  modulo the module docstring and __main__ guard.
• Bench page refreshed against the AMD EPYC-Milan host on a fresh
  machine fingerprint (37652a58e939). Speedups hold within VPS
  variance; absolute medians shift by ±3% versus the 0.0.3 snapshot.
  The dual-SHA LVIS caveat retires — every section now lives at the
  same SHA / fingerprint. The panoptic and synthetic-semantic cells
  exceeded the 5% relative-IQR gate (chronically noisy on this host —
  PNG decode dominates panoptic wall time, mmseg synthetic sits at the
  noise floor at 200-image scale); flagged inline with * per the
  renderer's existing convention.

Install vernier-cli 0.0.4

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/NoeFontana/vernier/releases/download/v0.0.4/vernier-cli-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/NoeFontana/vernier/releases/download/v0.0.4/vernier-cli-installer.ps1 | iex"

Download vernier-cli 0.0.4

File	Platform	Checksum
vernier-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
vernier-cli-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
vernier-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
vernier-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

0.0.3 — 2026-05-15

Release Notes

This is the diagnostic-surfaces and scenario-slicing release: instance
gains an oLRP error decomposition (Oksuz et al.), a detection-family
calibration summarizer (ECE / MCE / reliability), and a manifest-driven
slice-and-aggregate lane that runs one matching pass across N scenario
cells. Panoptic picks up boundary PQ. No paradigm shifts, no
crates.io additions — every kernel slots into the existing
vernier-core / vernier-panoptic / vernier-semantic surface.

Added

• LRP / oLRP error decomposition (ADR-0043, ADR-0044, ADR-0045) —
  Oksuz et al. (ECCV 2018 / TPAMI 2021) Localization Recall Precision
  as an opt-in metric alongside AP. vernier.instance.optimal_lrp(gt, dt, iou=Bbox()|Segm()|Boundary()|Keypoints()) decomposes detection
  performance into oLRP_Loc + oLRP_FP + oLRP_FN, minimised over a
  per-class confidence threshold tau. CLI gains --metric {ap,olrp}
  with ap preserving the existing headline-table contract. The Rust
  core lives in crates/vernier-core/src/lrp/; the ADR-0005 firewall
  is held (no edits to matching.rs / accumulate.rs / evaluate.rs).
  Pure-NumPy oracle is the correctness contract (ADR-0043);
  kemaloksuz/LRP-Error is an opt-in tripwire, not a parity gate.
  vernier.panoptic.optimal_lrp is a typed NotImplementedError stub
  — panoptic predictions carry no per-segment score so the tau sweep
  has nothing to scan; extension is a follow-up ADR.
• Boundary Panoptic Quality (ADR-0025 §Z1/Z2 amendment) —
  PanopticEvaluator(boundary=True, dilation_ratio=0.02) now ships
  under both parity_mode="strict" (bit-exact reproduction of
  bowenc0221/boundary-iou-api's coco_panoptic_api/evaluation.py
  at SHA 37d25586a677) and parity_mode="corrected" (deterministic,
  snapshot-based; segment-id-sorted iteration). Composition is
  iou = min(mask_iou, boundary_iou) — identical to the instance
  Boundary case (the prior Q3 row of boundary-iou-quirks.md had
  miscalled this; corrected in the same amendment). FN/FP attribution
  is unchanged; U6/U7/V1-V7/W1/W7 stand. The streaming runner threads
  boundary state per image with BoundaryScratch reuse, and
  distributed-eval partials hash the dilation_ratio into
  params_hash so silent boundary/instance partial mixing is rejected
  at envelope-validation time. No FORMAT_VERSION bump. Cityscapes
  panoptic (Z3) remains deferred.
• Detection-family calibration summarizer (ADR-0018) —
  ECE / MCE / reliability table for bbox / segm / boundary /
  keypoints. Opt-in via Evaluator.evaluate(..., calibration=True);
  the lazy result.calibration(iou=..., n_bins=15, binning="quantile", min_score=0.05, per_class=False, ...) re-fold
  returns a vernier.calibration.CalibrationResult (polars
  reliability / per_class plus scalar ece / mce). Re-folding
  with different params does not re-run matching. Streaming pairing:
  BackgroundEvaluator.finalize_with_cells() plus the
  vernier.calibration.StreamingSnapshot wrapper. Clean-room NumPy
  oracle is the correctness contract; 16/16 parity bit-equal at
  strict mode. Panoptic and semantic calibration are deferred
  (data-model prerequisites per the ADR's per-paradigm shape map).
• Slice-and-aggregate (ADR-0046) — manifest-driven scenario
  slicing across all three paradigms plus the vernier aggregate
  fan-in verb. Python:
  Evaluator.evaluate(..., manifest=..., cross_axes=...) accepts a
  dict, JSON / CSV path, or Arrow PyCapsule manifest and returns
  EvalResult.slices as a polars DataFrame (one row per
  (axis, value) cell). CLI: vernier eval --manifest weather.json [--cross weather,time_of_day] [--label NAME] [--metric {ap,olrp}]
  emits a v2 envelope; un-partitioned vernier eval keeps emitting
  v1 verbatim. vernier.aggregate(results, manifest, *, baseline=None, metric=None) and vernier aggregate result1.json result2.json --manifest runs.json --baseline clean fan N runs
  into a comparative table with <metric> (mPC) and
  <metric>__rpc (rPC) columns when --baseline is set. The
  tables= + manifest= cross product is a deliberate non-feature
  with a client-side recipe at
  docs/how-to/per-class-by-slice.md.
  New reference schemas: manifest-schema.md,
  aggregate-schema.md.

Install vernier-cli 0.0.3

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/NoeFontana/vernier/releases/download/v0.0.3/vernier-cli-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/NoeFontana/vernier/releases/download/v0.0.3/vernier-cli-installer.ps1 | iex"

Download vernier-cli 0.0.3

File	Platform	Checksum
vernier-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
vernier-cli-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
vernier-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
vernier-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

0.0.2 — 2026-05-12

Release Notes

This is the three-paradigm release: instance gains panoptic and semantic
siblings, distributed eval lands across all three, and the bench harness
brings real-model + alternatives numbers to the docs site. Two new
crates ship to crates.io (vernier-panoptic, vernier-semantic) plus
the vernier-partial leaf that holds the shared partial wire envelope.

Added

• Distributed-eval entry points on Evaluator (ADR-0035) — each
  paradigm's public Evaluator gains
  evaluate_to_partial(..., *, rank_id) -> bytes and a
  classmethod from_partials(...) -> Summary. Per-paradigm shapes:
  instance takes JSON bytes, semantic takes Dataset/Predictions,
  panoptic takes per-image tuples + categories= (the one
  asymmetry — PanopticDataset doesn't yet expose per-image
  accessors; closing that gap is a follow-up). The streaming
  substrate, the vernier-partial wire format, FORMAT_VERSION,
  partition-disjointness invariant, and the five paradigm-shared
  Partial* exception classes are all unchanged. The same DDP
  recipe works on instance, semantic, and panoptic.
• Distributed evaluation wire format (ADR-0031, ADR-0032) — new
  vernier-partial workspace crate holds the shared partial-envelope
  (magic + FORMAT_VERSION + framing + the five Partial* typed
  errors) used by all three paradigms. FORMAT_VERSION is a 1→2
  hard break (pre-1.0 policy). Cross-paradigm merge is structurally
  rejected (paradigm tag in the envelope). Determinism contract is
  paradigm-specific: instance preserves bit-exactness, semantic
  preserves it for any partition, panoptic only when the partition
  order matches the original GT order. BackgroundEvaluator reuses
  the same substrate via finalize_to_partial.

Changed

• Public-surface consolidation (ADR-0035, supersedes the public
  StreamingEvaluator portion of ADR-0013; amends ADR-0014, ADR-0031,
  ADR-0032). Each paradigm now exposes two classes: Evaluator
  (frozen config dataclass; batch + DDP entry points) and
  BackgroundEvaluator (in-training entry point; submit /
  finalize / finalize_with_tables / finalize_to_partial /
  context manager). The streaming pyclasses are removed from Python
  entirely; the Rust substrate stays and is reachable via new
  PyO3 functions (evaluate_*_to_partial, merge_*_partials) and
  via BackgroundEvaluator.

Removed

• vernier.{instance,panoptic,semantic}.StreamingEvaluator — the
  three streaming pyclasses are removed from Python entirely. They no
  longer appear on vernier._core, on any paradigm namespace, or
  under a vernier._impl shim. The Rust streaming substrate
  (vernier_core::stream::StreamingEvaluator<K>,
  StreamingPanopticEvaluator, StreamingSemanticEvaluator) remains
  as the implementation behind the new
  evaluate_*_to_partial / merge_*_partials PyO3 functions and
  BackgroundEvaluator's worker. No public deprecation shim — pre-1.0
  hard break.

• Evaluator.stream(...) factory on vernier.{panoptic,semantic} —
  removed alongside the public streaming class. Use
  BackgroundEvaluator(...) directly, or Evaluator.evaluate_to_partial
  / Evaluator.from_partials for DDP.

• StreamingEvaluator.snapshot(running=True) and its Rust-side
  snapshot_running() method — the biased fast path that ADR-0013
  itself flagged as inappropriate for quality gates.

• StreamingEvaluator.checkpoint() / restore() — these were
  NotImplemented thin wrappers around snapshot_to_partial /
  from_partials. The persistence story is now exclusively
  evaluate_to_partial → store bytes → from_partials on resume.

• BackgroundEvaluator.snapshot(), snapshot(peek=True),
  snapshot_with_tables(), and the non-finalize to_partial() on
  all three paradigms. Public surface is the consuming
  finalize / finalize_with_tables / finalize_to_partial only.

• BackgroundPanopticEvaluator.from_partials /
  BackgroundSemanticEvaluator.from_partials — vestigial (return-
  type bug carried them; no caller used them).

• Semantic-segmentation user docs (ADR-0028 PR-B10) — three new
  pages in docs/: migrate/from-mmsegmentation.md (semantic-side
  migration recipe with preset / streaming / NaN-vs-0.0 /
  binary-mask coverage), and explanation/three-paradigms.md (paradigm
  picker — when to reach for instance vs panoptic vs semantic, why
  they're sibling submodules rather than a single evaluator with a
  knob). README updated to
  feature the three-paradigm surface in a top-level section
  alongside the install commands; mkdocs.yml nav surfaces both
  new pages plus the previously-orphaned panoptic migration
  guide.

• Semantic-segmentation streaming evaluator (ADR-0028 PR-B9
  partial — streaming only; Breakdown / result-tables follow-ups
  scoped to a future PR). New
  vernier_semantic::StreamingSemanticEvaluator is a flat
  O(n_classes²) accumulator over ConfusionMatrix: update(image_id, gt, dt) folds via the same accumulate_confusion kernel the batch
  path uses; snapshot() is constant-time relative to image count
  (per ADR-0013, no fast-vs-running mode distinction needed). FFI
  pyclass vernier._core.StreamingSemanticEvaluator is registered
  on the module; the Python Evaluator.stream(n_classes, ignore_label=None) factory returns a fresh streaming evaluator
  carrying the parent's parity_mode. Load-bearing invariant
  (pinned by tests/python/test_semantic_streaming.py::test_streaming_finalize_bit_equals_batch_evaluate):
  streaming finalize() is bit-equal to batch evaluate(...) over
  the same images on f64 outputs. 10 new Python tests + 7 new Rust
  tests; total workspace 472 Rust + 376 Python tests pass.

• Semantic-segmentation Python wrapper + per-dataset presets
  (ADR-0028 PR-B5) — new vernier.semantic submodule (per ADR-0029)
  exposing Dataset / Predictions / Evaluator frozen dataclasses
  plus Summary / ClassSemanticStats / ConfusionMatrix
  re-exports of the FFI pyclasses (under their unprefixed names).
  Dataset.from_arrays and Predictions.from_arrays accept any
  unsigned-integer dtype; the wrapper preserves the input dtype and
  the FFI/kernel walks at native dtype (since ADR-0037).
  Dataset.from_files / Predictions.from_files decode single-
  channel PNG label maps via lazy-imported Pillow (raises a
  structured ImportError if Pillow is missing); RGB-encoded panoptic
  PNGs are rejected with a typed message pointing at
  vernier.panoptic.Dataset. Predictions.from_binary_masks
  implements the AN2 per-class binary-mask merge with explicit
  merge ∈ {"argmax", "first", "highest_class_id"} selector and
  unlabeled_class parameter (quirks AN3, AN4). Per-dataset
  presets Dataset.cityscapes / ade20k / pascal_voc bake the
  canonical (n_classes, ignore_label) constants from
  vernier_semantic::parity::*. 23 new Python tests cover the
  wrapper round-trip, dtype handling, ignore-label / label-remap
  propagation, binary-mask merge rules, RGB rejection, and
  end-to-end PNG decode + evaluate.

• Semantic-segmentation FFI surface (ADR-0028 PR-B4) —
  vernier._core.evaluate_semantic_from_arrays(gt_label_maps, dt_label_maps, n_classes, parity_mode, *, ignore_label=None, label_remap=None) is the load-bearing pyfunction that drives the
  Rust kernel + summarize pass under py.detach (ADR-0006). Inputs
  are dicts mapping image_id (int) → 2-D numpy.ndarray of dtype
  uint32. New pyclasses SemanticSummary, ClassSemanticStats,
  ConfusionMatrix expose the per-class and global metrics; the
  confusion matrix is materialized as a 2-D numpy.uint64 array
  via ConfusionMatrix.counts() (ADR-0028 §F1 first-class output).
  GT image-id ordering is sorted for deterministic accumulation
  (quirk AM5 aligned). label_remap is pre-applied to DT
  buffers at the FFI boundary (quirk AK2) so the hot kernel
  loop avoids per-pixel dict lookups. PNG-decode (from_files) and
  binary-mask (from_binary_masks) variants land in PR-B5
  alongside the per-dataset preset constructors that drive them.
  14 Python smoke tests pass; full workspace 465 Rust + 343 Python
  green.

• Semantic-segmentation kernel + summarize (ADR-0028 PR-B3) —
  vernier_semantic::kernel::accumulate_confusion per-image
  histogram fold (one pass over flattened (H, W) slices into a
  u64 (n_classes, n_classes) matrix; ignore-label mask before
  the bincount per quirk AJ2; out-of-range DT silent-skip per
  AI4 strict-MS path). ConfusionMatrix is a flat-Vec<u64>
  row-major shape that doubles as the FFI (N, N) numpy-view
  source. vernier_semantic::summarize::summarize derives the
  seven headline outputs (mIoU, FWIoU, pixel accuracy, mean
  accuracy, per-class IoU/accuracy/precision, plus the confusion
  matrix as a first-class output per AL8). parity_mode
  selects NaN vs. 0.0 for zero-support per-class entries (quirk
  AL2); means skip zero-support classes regardless of mode
  (AL3, mirroring panopticapi W2 and LVIS AB3). 16
  unit tests (kernel + summarize) on hand-computed fixtures, all
  pass in --release and debug. No SIMD per ADR-0028 §"Numerical
  layout" — the kernel is integer/memory-bandwidth bound. Dataset
  constructors and FFI surface land in PR-B5 / PR-B4 respectively.

• Semantic-segmentation crate scaffold (ADR-0028 PR-B2) — new
  workspace member crates/vernier-semantic/ with Cargo.toml /
  lib.rs / error.rs / parity.rs. Re-exports
  vernier_core::parity::ParityMode per ADR-0028 §"Workspace and
  dependency direction" — the first dep-edge asymmetry vs.
  vernier-panoptic ⊥ vernier-core, justified by concrete reuse.
  Pins the per-dataset ignore-label conventions
  (CITYSCAPES_IGNORE_LABEL=255, ADE20K_IGNORE_LABEL=0,
  PASCAL_VOC_IGNORE_LABEL=255), class counts, and
  SEMANTIC_PARITY_EPS placeholder. SemanticError enum surfaces
  the corrected-disposition...