Greg v1.7.0b1

Test build of 1.7.0. Not for general use.

What it will contain when it ships:

Added

• The escalation ladder. The three thresholds are one track with three
  handles in Greg's panel, in the mood colours the card already uses. The
  handles cannot cross, so chaos sitting below medium, which made him skip a
  mood for no visible reason, is no longer something that can be set.
• Timing in the panel too, as three numbers you type over rather than
  three more full-width sliders: how long until he calms down, how long until
  he notices the silence, and how often the existential crises come round.
• All six are still in advanced settings as well. Nothing has moved out from
  under anybody, and if they end up crossed there Greg puts them in order.
• Line backups. Export to a file writes everything you have written to
  greg-lines-<date>.txt. Import from a file reads one back, asking first
  whether to add the file's lines to yours or replace yours with them, and
  saying how many lines a replace would throw away.
• The file is plain text: ## en, ### soft, and a line per bullet. Anything
  else in it is ignored, so you can also write a batch of lines in a text
  editor and import them, and a file somebody sends you can be pasted straight
  into a GitHub issue.
• greg.import_lines does the same from an automation. One call rather than
  one per category, so a restore is a single reload.
• greg.set_options now accepts the thresholds and the timing values.

Fixed

• Custom Piper voices were ignored. A .onnx installed by hand is not in
  the list Wyoming advertises, and Greg refused any name he could not find in
  that list, so a custom voice never played while the same name worked fine in
  an automation. He now asks for it and notes in the log that he did not
  recognise it. Reported by
  rujelus22 on the forum
  thread.
• An en_US voice was checked against the wrong locale. Greg's codes are
  bare, so English fell through to whichever locale sorted first, usually
  en_GB, and an en_US voice was then not recognised. A voice now decides
  the locale, as long as the engine offers it and it speaks the language Greg
  is speaking. Same report.
• A voice name that is right apart from its accents or its capitals is now
  corrected to the engine's own spelling rather than refused, which is the
  pt_PT-tugao-medium trap from 1.5.5 fixed rather than merely explained.
• The existential interval defaulted to 37 minutes in one file and 42 in
  another, so a fresh install that never opened the advanced dialog ran five
  minutes fast.
• Thresholds stored in the wrong order, which the dialog has always allowed,
  are put back in order wherever they are read, so an install that has been
  quietly skipping a mood for months stops doing it.

Internal

• Tests live in the repo now, and run in CI: python -m unittest discover -s tests, plus node tests/ladder.test.mjs and node tests/backup.test.mjs,
  which run the real panel in a stubbed DOM. No pytest, no Home Assistant.
• tests/preview.html opens Greg's panel against a fake Home Assistant, for
  looking at a change without an install to hand.