Release 1.1.0: TVAE KL weight, bounded spawner history, RiteWeight reweighting

[namus]

Three additions and one fix, all default-compatible. Every new setting defaults to
what the code already did, so existing configuration files and published results are
unaffected.

Added

spawning.history_window — how much of the campaign a spawner scores against.
Default (null) is unchanged: the whole history. An integer gives PaCS-MD-style
cycle-local selection; 0 scores only the current iteration. A coverage objective
is only definable over cumulative history, and this makes that claim testable under
an otherwise identical loop.

adaptive_model.tvae_beta — the KL weight in loss = mse + beta * kld / n_features.
The estimator was previously built without this argument, so every run silently used
deeptime's default of 1.0 and no user could change it; answering "which beta did you
use?" meant reading a dependency's source. Default stays 1.0; checkpoints written
before this release restore to 1.0 rather than to a new default.

trails_md.analysis.riteweight — randomized iterative trajectory reweighting
(Kania et al., PNAS 123, e2529246123, 2026): a stationary distribution from
adaptively-sampled data without a lag time and without assuming cluster-level
Markovianity, complementing the MSM route. Independent implementation from the
published algorithm; the authors' reference code carries no licence statement and was
deliberately not copied or vendored.

Fixed

• The shipped template advertised encoder_hidden_dims: [64, 32] / decoder_hidden_dims: [32, 64]
  while the built-in defaults are [256, 128] / [128, 256], so copying the template
  trained a different network from omitting the block. The template now states the
  real defaults.
• trails_md.analysis exported only data, leaving riteweight importable by full
  module path alone.

Documentation

• New docs/reweighting.md: MSM vs RiteWeight and how each fails, how to build segment
  pairs without crossing a respawn, and the limitation both share — neither fixes
  mis-coverage.
• docs/cv_methods.md: the TVAE loss in full, beta and its per-feature
  normalisation, lagtime counted in frames (physical lag = lagtime × stride × dt),
  and an explicit statement that time-lagged pairs never span a respawn.
• docs/configuration.md: history_window, tvae_beta, dropout_rate, decoder widths,
  SPIB hyperparameters, and a section on cumulative vs cycle-local selection — including
  that the comparison must be made against aggregate simulation time, not wall clock.
• .zenodo.json so the archived release carries proper metadata.

Tests

260 passing. 27 new: window semantics (including that a window of 0 must not slice to
the whole list), beta reaching the estimator and surviving a checkpoint round trip,
and RiteWeight's cluster-count independence plus recovery of structure finer than its
own clustering.