Skip to content

Packvium 1.7.0

Choose a tag to compare

@github-actions github-actions released this 18 Oct 16:17
· 4 commits to main since this release
6dce786

The reference validator ships in the crate, quality looks ahead in its block solver and
finishes inside its time limit, and extreme points scan the candidate set they were meant to.
No request or result field changes.

Added

  • validate_json(request, result) and validate_json_outcome, with ResultIssue,
    MalformedResult { kind, message } and RefusalKind: the reference validator over the JSON
    text, returning the same issues, code and detail, in the same order, as every other package.
    It shares no code with the solvers or IndependentValidator, which is unchanged.
  • PlacementConstraint::distinguishes_instances, a defaulted trait method, true unless
    overridden, so existing implementations compile and keep their behaviour. A constraint that
    reads only item attributes may return false to let the portfolio drop a start order that
    repeats another's item-type sequence.

Changed

  • A CandidateScorer leads the extreme-point rank. The rank is the tuple (sum of
    scorer scores, far corner z, y, x, origin, rotation); before, a scorer's score was added to
    one weighted sum of the far-corner ticks at the weight of one x tick. A scorer that returned
    small nudges meant as tie-breaks now decides the rank; return 0 where it has no preference.
    The trait's signature is unchanged, and no request without a custom scorer answers
    differently on the measured corpora.
  • Every schema enum value is refused as not_allowed. A value outside its list in
    allowed_rotations, the output and dimensional-weight units, shape_type,
    ground_contact_rule or configuration.solvers is invalid_request with reason not_allowed
    at its JSON Pointer. An unknown output unit was copied through; ticks as an output unit is
    now refused
    , as the schema always said.
  • quality looks ahead in the block solver, completing its leading blocks greedily
    and committing the best, so the answer is never worse than the greedy pass. On our benchmark
    85 of 89 answers improved and none got worse. Under a tight effort budget the block start can
    now report truncated / effort_limit.
  • quality finishes inside its limit. The staged beam refines the best greedy
    order instead of the first two, and the spatial index stops at the first hit without
    allocating. On the 22 pinned benchmark runs 1.6.0 needed more than 20 s of CPU on 10; this
    release needs at most 8.1 s.
  • Extreme points scan the exposed corners and their projections, not every
    combination of placed faces. With solvers: ["extreme_points"] alone, over 67 benchmark
    instances, fast/balanced/quality leave 2711/1477/1297 items out instead of
    3658/3048/2643. The full portfolio's answers did not change on the 89 benchmark requests, and
    its quality CPU fell to 42%.
  • Portfolio starts go to distinct item orders. A start that repeats another's
    item-type sequence runs a neighbouring order instead, kept only on a strict win. A stacking
    scene that took two containers now takes one.
  • no_eligible_container is reported proven, and insufficient_support is reported
    (observed) where the support rule is the only blocker.
  • A nesting_height not below the item's height is refused as invalid_value at its
    pointer, and with no output block the result is written in units.length's schema code.

Fixed

  • incompatible_tags is applied item against item, not against container tags, as the
    validator and every other engine apply it.
  • A panicking start fails a parallel portfolio as it fails a sequential one; it was dropped.
  • Axle-balanced candidate points use the envelope length under clearance and are not offered
    twice.