Changes in version 0.0.0.9001                      

  - Reject stale temporal partitions, incompatible radiometric
    prediction units, invalid survival inputs and misaligned named
    matrix outcomes. Reconstruction evaluation requires an explicit
    reference in the fitted output space. Refit legacy models lacking
    the new input-unit contract.
  - Preserve selected-frame observations, tri-state masks and complete
    band metadata through hyperspectR conversion; reject stale content
    identities and hidden live state in persistent metadata.
  - Align backend availability checks with startup integrity validation
    and stop owned workers after malformed responses. Ignore external
    pip configuration during explicitly requested environment setup.
  - Update the CPU reference profile to Torch 2.13.0, NumPy 2.2.6 and
    setuptools 84.1.0. Retire affected generated profiles and block
    automatic vLLM setup until a replacement is reviewed; existing
    runtime files remain preserved.
  - Honor the workbench's 50 MiB upload limit, bind setup consent to the
    reviewed resolution, and show unsupported-platform restrictions
    before job submission.
  - Make development installation and first use executable from a fresh
    session, repair packaged-document links, pin shared lint
    configuration and exclude Python bytecode caches from source
    distributions.

                     Changes in version 0.0.0.9000                      

  - Standardize the named v5 package logo across README, documentation
    pages and available application headers, with matching light and
    dark artwork.

  - Added formal spectral/sequence data contracts, bounded access and
    explicit companion-package source bridges with synthetic conformance
    fixtures.

  - Added grouped workflows, training-only recipes,
    statistical/chemometric and conventional model engines, nested
    tuning, explicit evaluation and conditional whole-unit uncertainty.

  - Added immutable run readback, runtime metadata resolution, reviewed
    isolated installation, owned workers and explicit named upstream
    admission contracts.

  - Added separate trainable neural/temporal reference implementations;
    named pretrained models, instruments and scientific transfer remain
    unqualified.

  - Added a local Shiny workbench, eleven executable development
    articles and the shared themakR website template. Browser/platform
    acceptance remains separate.