fix(hooks): stop the guard from blocking heredocs and unindexed sessions (v2.1.2)

[denfry]

Problem

The PreToolUse guard from 2.1.1 blocked calls it should have let through. In a real session it:

• kept intercepting after the agent had already run codebase-index search, because that search was the one that built the index: the guard saw no index, returned early, and never recorded that the index was used;
• denied a Bash call that only wrote a PowerShell script via cat > x.ps1 <<'EOF', because ; Select-String "BUILD FAILED" inside the heredoc body looked like a code search;
• denied the retry too, since the retry was keyed by the full tool input (description included) and the regenerated command differed by a few characters;
• cut grep "a\|b" at the \|, producing a broken suggested query.

Solution

• Session state moves to the system temp directory (CBX_HOOK_STATE overrides it) and is written before the index lookup, so the first index command always turns the guard off.
• Heredoc bodies are stripped before scanning; newlines separate commands.
• Retries are matched by the normalised search term.
• Quoted arguments are parsed whole.

Verification

• 6 new tests in tests/test_hooks_guard.py, all failing before the fix; full suite passes (coverage 86%), ruff and mypy clean.
• Replayed every Bash call of the affected session through the new guard: none denied. In a fresh session a real grep --include=*.java is nudged once and its retry passes; the heredoc scripts pass.

Compatibility

Sessions started before the upgrade keep their old state inside the index, so they may be nudged once more. No settings changes needed.

Also bumps the version to 2.1.2 with every mirror synced.