fix: make the hook cross-platform by removing the bash wrapper

[jesse-quinn]

Fixes #18

Problem

hooks/hooks.json invoked scripts/prompt-guard-wrapper.sh, a bash-only
script that logged to /tmp with Unix paths. On native Windows (no
WSL/git-bash) the hook fails, which depending on how the exit code is
interpreted either breaks prompting or silently disables protection.

Fix

Eliminate the bash layer entirely and invoke Node directly.

• hooks/hooks.json now runs node ${CLAUDE_PLUGIN_ROOT}/scripts/prompt-guard.js
  (same ${CLAUDE_PLUGIN_ROOT} interpolation the plugin already used; 10s
  timeout unchanged).
• scripts/prompt-guard.js absorbs the wrapper's only real job -
  opt-in debug logging. When PRIVACY_GUARD_DEBUG=1, it writes execution
  metadata (timestamp, CLAUDE_PLUGIN_ROOT, cwd, Node version, exit
  path/code) to debug.log under the platform cache dir, resolved
  cross-platform by a small os.platform() helper:
  • POSIX: XDG_CACHE_HOME or ~/.cache (unchanged from before)
  • Windows: %LOCALAPPDATA%
    Matched secret/PII values are never written to the log.
• Deleted scripts/prompt-guard-wrapper.sh.
• README debug section updated with the Windows log path.

Behavior parity

• Block-decision JSON still goes to stdout; process exits 0 on both
  allow and block (a non-zero exit would make the hook protocol treat the
  block as a non-blocking error and let the prompt through).
• The wrapper discarded stderr when debug was off. Without the wrapper,
  stderr would now reach Claude Code, so internal errors are routed to the
  debug log (when enabled) and stderr stays quiet otherwise. Error paths
  still exit non-zero, matching the wrapper's pass-through of a failing exit
  code (a non-blocking error - the prompt is not silently altered).

dist/ note

These changes touch only scripts/ and hooks/ (and README.md) - no
src/ changes - so dist/ does not change. git status dist/ is clean,
satisfying the CONTRIBUTING requirement that dist/ match src/.

Manual verification (macOS, Node v26.5.0)

Secret prompt - prints block JSON, exits 0:

$ echo "my api key is sk-proj-abc123xyz1234567890" | node scripts/prompt-guard.js
{
  "decision": "block",
  "reason": "...Found 1 sensitive item(s):\n  - OpenAI API Key: sk-p…7890 ..."
}
EXIT=0

Benign prompt - no output, exits 0:

$ echo "what is the weather like today" | node scripts/prompt-guard.js
(no output)
EXIT=0

PRIVACY_GUARD_DEBUG=1 - creates the log with metadata only (no secret):

=== Hook Execution 2026-07-19T02:34:00.057Z ===
CLAUDE_PLUGIN_ROOT: /fake/plugin/root
CWD: /.../claude-code-privacy-guard
Node version: v26.5.0
Exit path: block
Exit code: 0

Grep of the log for the secret value (sk-proj, abc123, 7890) - no match.

npm install && npm run build && npm test: all 159 tests pass; dist/
unchanged.