fix(hook-kit): reconcile the three diverged bash-guard copies into one union

[DrumRobot]

Problem

bash-guard.py was living as three independently edited copies. Fixes landed on whichever copy happened to be executing, and never travelled back to git — so the tracked file was missing six functions that were actively guarding real sessions.

Copy	Fns	Has	Lacks	Self-tests
tracked (staging branch)	11	staging base gate	force-push fn, heredoc stripping	93
plugin cache (es6kr)	13	heredoc stripping, worktree-add exemption, pr-merge-ready check, transcript-aware pr-create-draft	staging gate	99
~/.agents/skills copy	14	conditional force-push guard + 2 helpers	the cache-only trio	93 (none for its own guard)

The runtime loads from the plugin cache, not the marketplace — confirmed from installed_plugins.json's installPath, not inferred from error-message paths. Two plugins (es6kr, task) cache the same hook, so both run, and they disagree.

Change

Union merge. The cache copy is the skeleton because its two overlapping functions are strict supersets of the tracked ones; the staging gate comes from the tracked copy and the force-push trio from the third. Three module constants and one transitive dependency came along with them.

Behaviour change worth review: force-pushing stops being an unconditional block. Targeting a protected branch is still denied; a lease-checked push to a feature branch is now allowed. That conditional-allow fix existed only in the untracked copy until now.

Test plan

• --test self-suite: 105 cases pass (93 tracked + 6 cache-only + 6 staging)
• Behaviour matrix returns the intended verdict for: protected-branch force-push (deny), lease-checked feature-branch push (allow), plain push (allow), staging base without a divergence check (deny), same with the check (allow), hard reset (deny), pr-create without draft (deny)
• python3 -m py_compile
• make lint, make verify-spec, make test (9 passed)
• Reviewer confirms the force-push relaxation matches intended policy
• Reviewer confirms no guard regressed relative to the copy they rely on

Known gap (not in this PR)

The copies will drift again: git tracks the marketplace source while the runtime reads the cache, and the sync target is two cache directories rather than one. A structural fix (symlink, or a copy-comparison step in the hook audit) is a separate change. The other staging branch already carries feat(hook-kit): check the runtime copies a hook actually has, which looks like the start of that work.

While committing this, the task cache copy — which lacks the heredoc stripping the es6kr copy has — denied the commit by matching a phrase inside the message body. That false positive is the drift's cost, observed rather than hypothesised.