[Review only] Align datasets to a common origin

[tjlane]

! This PR was vibe-coded.

Important

This is a review-only aggregate view of the existing stacked PRs. It adds no new commits and should not be merged directly. The smaller PRs remain the merge path.

Stack context

This draft collects the common-origin layers of #31 in one diff:

• constrained origin Fourier grid: Calculate phase agreement over symmetry-allowed origin shifts #355
• bounded-memory FFT search: Search large origin-shift grids with limited memory #356
• symmetry-aware origin ranking: Refine and rank possible origin shifts #357
• phase matching and origin application: Match phase data and apply a chosen origin shift #358
• public DataSet orchestration: Automatically reindex and align phases to a common origin #359
• weighting and optional hand inversion: Support phase weights and optional hand inversion #360
• confidence gates and final public contract: Warn about weak alignments and reject ambiguous ones #349

It is based on the completed reindexing branch (#351–#354), because align_phases() must attempt correlation-based reindexing before origin alignment whenever the dataset permits it. The origin-ambiguity predicate itself is introduced in the shared foundation PR #351.

What this implements

• determines and searches only symmetry-allowed origin translations
• represents continuous polar freedom with an exact integer basis and quotients trivial centering translations
• evaluates each discrete origin coset with an intrinsic-dimensional FFT
• switches to a bounded-memory slab search for large grids and refines the strongest sampled maxima
• matches phases by HKL, supports amplitude/FOM or uniform weighting, and transforms phase and complex columns on a copied DataSet
• automatically reindexes before origin alignment when permitted; hand inversion remains opt-in
• reports Phenix-sign origin shifts, candidate correlations, runner-up scores, and correlation gaps
• warns for weak correlations and rejects low or non-unique solutions

Reviewer focus

• derivation of the symmetry-allowed cosets and centering equivalence
• Fourier sign, integer polar basis, periodic maxima, and the bounded-memory top-K policy
• Phenix sign convention and preservation of row order/MTZ dtypes
• default amplitude-product weighting, FOM validation, and opt-in hand inversion
• interaction between automatic reindexing and origin alignment
• empirical warning/error thresholds and the no-clear-solution gates

Please leave line-level review on the constituent PR where possible; this draft is intended to make the complete common-origin story easy to inspect in one place.

Related: #31, #174