v0.3.0: chore: add prepare script so git-hosted installs build dist-js

Driven by transcript mining of real agent sessions: 80% of all tool calls
were execute_js, and most of those were hand-rolled variants of four
patterns. These are now first-class tools.

Added

• read_text tool — structured text scraping: per-element
  {tag, text, visible, attrs?} for a CSS selector, whitespace-collapsed,
  hard-capped output with a truncated flag, optional scope_selector.
  Replaces querySelectorAll(...).map(el => el.textContent) execute_js snippets.
• inspect_element tool — bounding rect, computed styles (default
  layout/color set or explicit style_props), classList and attributes for
  visual/layout QA. Replaces getComputedStyle/getBoundingClientRect snippets.
• dispatch_pointer tool — synthetic pointer/mouse gesture chains
  (click, dblclick, down, up, hover, drag) dispatched on the
  exact matched element with no interactive-ancestor retargeting. Supports
  element-relative offset (canvas hit-testing), absolute or relative drag
  destinations, interpolated drag moves double-dispatched on the element and
  document (d3-drag compatible; synthetic events bypass setPointerCapture),
  buttons and modifier keys.
• app_bridge tool + window.__MCP_BRIDGE__ registry — host apps
  register named (optionally async) helpers via
  __MCP_BRIDGE__.register(name, fn, description); agents discover them with
  action='list' and invoke with action='call'. Results are serialized and
  truncated at max_chars. The sanctioned path for store snapshots and
  app-level actions instead of execute_js against app internals.
• click / query_page(find_element) sharpening — new optional
  scope_selector, match (exact|contains), and nth parameters for
  disambiguating repeated text/selectors within a container.

Changed

• find_element text matching collects all candidates, prefers innermost
  matches hoisted to their nearest interactive matching ancestor, and ranks
  interactive > semantic > shortest text — it no longer returns a huge
  container div whose textContent happens to include the target string.

Fixed

• Linux screenshot crash on windows with non-ASCII titles: bumped xcap
  0.0.4 → 0.9 and image 0.24 → 0.25 (JPEG encoding now converts RGBA → RGB,
  required by image 0.25). Ported from #22 — thanks @leiha.