v0.1.1 — Correctness & safety hardening

Sediment v0.1.1 — Correctness & safety hardening

A focused pass after a full adversarial review of the analyzer. There are no new
commands; the analyzer is simply more conservative and more correct. Every change
errs toward under-claiming rather than guessing — a key is reported as owned,
cleaned, or safe to delete only when the source genuinely proves it.

Fixed

• Cross-namespace resolution. Names resolve to their fully-qualified form, so
  two classes or functions that share a short name in different namespaces no
  longer cross-resolve, and a same-named method in an unrelated class can no
  longer credit an uninstall callback.
• Local-variable binding. Variables bound by function parameters, foreach,
  global, static, catch, and destructuring now resolve to dynamic; a
  single literal assignment can no longer claim a variable whose value varies at
  runtime and flow into a generated deletion.
• Anonymous classes are keyed per file, closing a cross-file symbol leak.
• Table SQL is read with one anchored, statement-aware parser shared by
  detection and cleanup: CREATE TABLE inside an INSERT value is ignored, and
  every statement in a multi-statement dbDelta is captured.
• Cleanup scoping. Only confident removals count, and only inside the
  plugin-root uninstall.php (a top-level statement or a function it invokes) or
  a registered callback matched case-insensitively. A dead function in
  uninstall.php, a $wpdb->prepare(...)-only drop, and a pattern/dynamic
  removal no longer credit cleanup.
• Never crash. A scan survives an unreadable directory, and the first pass is
  guarded like the second.

Added

• The WordPress core allowlist now covers roles (wp_user_roles and friends),
  uninstall_plugins, recently_activated, and more core cron hooks — so a
  generated uninstall.php can never emit a delete for them.

Changed

• The generator rebuilds the {prefix} token from $wpdb->prefix for options,
  cron, and transients (not only tables), wherever it appears in a key.
• The grader counts unique keys rather than call sites, treats unknown-autoload
  options as autoloaded, caps the score when there is no uninstall routine, and
  reports low coverage instead of "creates no data" for an all-dynamic plugin.
• Scanned source is escaped before it reaches the terminal report.

Full details are in the changelog.