v0.3.1a1

First alpha of the v0.3.1 line. Two threads of work: a code-review
sweep that tightens the integration's robustness without changing
user-visible behavior, and two UX fixes for the rule editor that
came out of dogfooding the v0.3.0 GA.

What's new

• Panel YAML view: Save / "Show visual editor" are visible again
  on long configs. ha-code-editor's host max-height doesn't
  constrain the internal CodeMirror surface — the editor grew with
  content and pushed the action row off-screen. Switched to
  --code-mirror-max-height (the CSS variable HA's own
  automation editor uses), which makes the editor scroll
  internally and keeps the action row visible. Same fix applied
  to the per-rule code-mode editor.

• Rule editor: backdrop click and ESC no longer silently throw
  away unsaved edits. The editor now tracks a dirty flag; when
  the user has unsaved work and clicks off the dialog (or presses
  ESC), a "Discard changes? [Keep editing] [Discard]" confirm
  appears. Save success and clean dismissal pass through
  unchanged. Implementation note: HA's modern ha-dialog wraps
  a native <dialog> and binds pointerdown (not click) for
  backdrop close, so the guard is a document-capture
  pointerdown listener — bubble-phase listeners on the element
  are too late.

Internal hardening (from code-review sweep)

• Validation size caps: mapping at 200 entries, thresholds at
  50, source_attribute length at 255, replace_all batch at
  1000. Bounds the synchronous validation/scan loops against
  pathologically large payloads.
• Validation tightening: empty thresholds lists explicitly
  rejected; non-string mapping keys rejected (YAML's bare 1:
  was silently coerced via str(), producing rules that didn't
  fire); non-string source_attribute rejected (used to be
  vol.Any(str, None) which accepted weird types via voluptuous
  coercion).
• Setup race fix: hass.data[DOMAIN] and the store/injector
  pointers are now wired before injector.async_start() and WS
  command registration, so a WS client racing setup can't see a
  partially-populated hass.data.
• Injector attr pre-check: state-update attribute writes only
  allocate a new dict when at least one of color / icon /
  background actually changed. Eliminates a per-event allocation
  for the common no-op case.
• Painter memory hygiene: light-DOM and shadow-DOM
  removedNodes from the MutationObserver now prune
  knownHosts immediately, instead of letting the set grow
  until the next full sweep. Fixes a slow leak in busy
  Lovelace dashboards with cards rotating in and out.
• State watcher: repaintAll() now only fires when the
  smart_icons: decoration attrs were involved (either present
  on the old state or present on the new). Filters out the long
  tail of unrelated entity updates.
• Frontend glob matching parity: matchGlob now translates
  fnmatch's [!abc] negation to JS regex [^abc], matching the
  backend fnmatch.translate semantics. Previously a glob
  pattern with [! rendered targets visually mismatched between
  the table and the live injector.
• Frontend source-default parity: serialize matches backend's
  rule of "only a pure single-literal target defaults source to
  the target." Multi-target rules with blank source no longer
  collapse to a single-source rule on save.
• Version-from-manifest: INTEGRATION_VERSION is now read
  from manifest.json at import time instead of duplicated as a
  hardcoded string in websocket_api.py. A regression test
  locks the two together.

Internal

• Build-time banner in the panel bundle ([smart-icons] panel bundle build <timestamp>) prints to the browser console on
  mount, so future "did my rebuild land?" investigations end with
  one console line instead of a guess.

Full changelog: v0.3.0...v0.3.1a1