v0.2.0 — anytime cascade across all four families

[v0.2.0] - 2026-06-03

Anytime cascade pattern now covers all four classical search families.
Under mode="auto" (the default), a family-appropriate meta-algorithm wins
selection and runs a cascade tuned to the problem shape — always returning
the best incumbent within budget instead of not_found.

Features

• AnytimeCSP — cascade [MinConflicts (only if @evaluate), Backtracking]
  wins selection under mode="auto" on CSP-shaped spaces. First phase to
  return a consistent complete assignment wins.
• AnytimeLocal — cascade [HillClimbing, SimulatedAnnealing, TabuSearch]
  wins selection under mode="auto" on pure-optimization spaces
  (@successors + @evaluate, no @goal). Lowest-cost incumbent wins across
  phases.
• AnytimeAdversarial — iterative deepening from depth 1 to max_depth
  over AlphaBeta (2-player) or Negamax (3+ player), threading the
  previous depth's principal variation as pv_hint into the next phase
  for move ordering. Wins selection under mode="auto" on adversarial
  spaces.
• Adversarial cancel-token cooperation — AlphaBeta, Minimax,
  Negamax check the token at the top of their recursion ((nan, None)
  sentinel that collapses to not_found at the root); MCTS checks it
  at the top of its iteration loop and returns best-so-far from the
  partial tree.
• PV-first move ordering for AlphaBeta and Negamax —
  AnytimeAdversarial feeds the previous iteration's principal variation
  in as a hint; α-β pruning becomes substantially more effective.
• Path population for Minimax, AlphaBeta, Negamax — they now
  return SearchResult.path as [(action, state), ...] like the other
  search families (contract change documented in the design spec).
• Mode-contract overrides for MCTS and Negamax — MCTS.score_for bumps
  to 53 under mode="approximate" (closing the prior gap where approximate
  did nothing for game spaces); Negamax.score_for bumps to 52 when
  space._players > 2 (closing a bug where AlphaBeta silently won
  "exact" selection on multi-player games despite not honouring
  _players).

Fixes

• CI: publish.yml renamed to publish.yaml to match the PyPI trusted
  publisher configuration.
• CI: tomllib.load(file) instead of tomllib.loads(bytes) for the
  version-consistency check.

Docs

• Modes & Anytime delivery guide now documents the four-family cascade
  table.
• Cancel-token protocol guide moves the adversarial family from "watchdog
  backstop" to "cooperates"; only IDAstar and CSP recursion now rely
  on the watchdog.
• API reference exposes the four meta-algorithms (AnytimeAStar,
  AnytimeLocal, AnytimeCSP, AnytimeAdversarial).
• README and docs/index.md algorithm-family tables call out the default
  meta-algorithm per family.