v1.2.0

Auto-Explore now ranks with a model that was trained on this cellular automaton, instead of one
borrowed from photographs. The optional CLIP path is gone — code, worker, dependencies, settings,
and all.

Added

• A HexLife-native interest model ranks Auto-Explore, and it is the default. Statistics still do
  the cheap screening and the hard kills, so a dead or saturated candidate is rejected without any
  model involvement. Every survivor of the confirmation burst then gets an exact 32-frame trajectory
  scored by a 21.5k-parameter model that runs from public/models/hexlife-interest/ inside your own
  browser — no third-party host, no CDN, nothing downloaded at runtime. The same pass returns a
  32-value behavior descriptor that drives novelty pressure and the illumination archive.
• A displayed score that means the same thing tomorrow. The model's raw preference utility has no
  meaningful absolute scale, so the shipped artifact carries a calibration that maps it to its
  percentile against a frozen reference corpus. The transform is monotonic, so rankings are
  unchanged, but a 0.87 in one session is the same 0.87 in the next.
• "Statistical only". One radio button restores the deterministic, model-free objective exactly
  as it behaved before this release — same populations, same champions, same reset seeds, same banked
  scores, pinned by a golden test.

Changed

• Auto-Explore is organised into tabs. Run status and Start / Pause / Stop / Stop & Keep stay
  visible at all times; Setup, Objective, and Finds hold everything else. The tabs are
  keyboard-navigable (arrows, Home, End), remember which one you were on, and behave the same on
  mobile Discover as in the desktop panel. Owner-facing Export selected and Evaluate selected
  moved into a collapsed Model Tools area under Objective.
• The curated library grew from 39 to 73 rulesets. Public and personal entries now share one
  source-filtered list, and duplicate personal copies of built-in rules can be removed safely.
• The keyboard shortcut viewer now uses the space it is given. Resizing its desktop panel grows
  the keyboard in both directions; modifier layers now distinguish Ctrl from Ctrl+Shift; search
  spans keys, actions, categories, and every layer; and the categorized fallback is available from
  mobile More Options. V remains the Flat/3D Torus toggle, while recording moves to Shift+V
  (start/stop) and Ctrl+Shift+V (pause/resume) to remove the previous duplicate binding.
• Torus-aware capture and re-centering. Capture Studio now exports the active 3D projection, and
  hold-H wrap-around shifting works without losing the orbit strategy.
• Ruleset relatives and family naming. Saving a ruleset now shows its nearest named relatives in
  the strictest shared constraint space and can continue a family name with Roman-numeral suffixes.

Removed

• The optional CLIP perceptual objective, entirely. The embedding service and its worker, the
  perceptual novelty and archive modules, the cell rasterizer, the noise-prompt contrast battery, the
  CLIP model picker, the perceptual status and progress UI, the openEndedness scoring term, the
  "Maximal Novelty" preset, and the embedding-based tag suggestions are all deleted. Because that
  term was always dropped-and-renormalized when absent, statistical scores are unchanged.
• Text-prompt target search. It only ever worked because CLIP had a text encoder; the native
  model has none, so a supervised prompt search would have been a control that could not work.

Fixed

• A run can no longer be stalled by the model. Inference has a deadline, an abort signal wired to
  the run token, and a fallback to the confirmed statistical score. A failed load, a timeout, a bad
  descriptor, or a mid-flight Stop all degrade to statistics instead of leaving a generation pending.

Notes for upgraders

• Nothing about worlds changed: ruleset codes, HXW1.… world codes, share links,
  <hexlife-world> attributes, and determinism are all untouched. Old share links and exported find
  packs still decode; fields that only meant something to the CLIP path are ignored on import.
• The old perceptual archive in local storage is discarded on first load and replaced by a
  native-descriptor archive that self-invalidates when the model changes. Your gallery of finds is
  not affected.
• The shipped model is labelled beta, not accepted. On the project's curated sanity panel it
  orders interesting-vs-boring pairs correctly 67% of the time against the statistical objective's
  44%, but it has not cleared the strict corpus gates (direct owner hard-pair votes, locked test
  cases, mixed-grid corpus coverage, quantization), and the manifest says so.