fix(signing): register active signing key in allowed_signers so verify-commit passes

[dollspace-gay]

Summary

Closes GH#585. In a driver workspace, crosslink trust approve <agent-id> registered only the agent's key in allowed_signers. The driver's own SSH signing key — which configure_signing correctly set as the cache worktree's user.signingkey — was never added. Every signed hub commit from that workspace was therefore correctly signed by the driver's key but failed git verify-commit because the verifier didn't trust that key. crosslink sync immediately reported "N invalid signature(s) in last N commit(s)" after the very first trust approve. The natural diagnostic (crosslink integrity sign-backfill) reported "no unsigned entries — nothing to do" because the commits ARE signed, just not trusted.

Fix (Option B from the issue — the author's recommendation)

Self-healing invariant in configure_signing. After any of its three branches (agent worktree, driver workspace, driver-fallback-to-agent) selects a key and calls signing::configure_git_ssh_signing, it now also calls register_active_key_as_trusted which:

1. Reads the <key>.pub companion of the just-selected private key.
2. Loads allowed_signers and dedupes via the new AllowedSigners::contains_key (compares by <type> <base64> body — matches across different principals and trailing comments).
3. Idempotent: returns early when the key is already trusted under any principal.
4. Appends a new entry using the agent.json identity as the principal (<agent_id>@crosslink, mirroring trust approve's format), falling back to driver@crosslink when no agent.json exists.
5. Commits unsigned — chicken-and-egg: the new key may not yet be in any earlier commit's allowed_signers view, so signing this commit would re-trigger the very verify-commit failure we're fixing. Same navigation as publish_agent_key.

Why Option B over A (which was the smaller-diff alternative — just patch trust approve): Option B covers both the fresh-bootstrap case AND the driver-key-rotation / re-init case the author flagged as B's structural advantage. Every signing-key selection self-heals.

Test plan

• cargo fmt --check — clean
• cargo clippy --lib --bin crosslink --tests — clean on touched files (pre-existing tests.rs warnings at lines 1086/1087/2538/2548 are unrelated)
• cargo test --bin crosslink — 2850 passed (2842 baseline + 8 new)
• 8 new tests: 5 for AllowedSigners::contains_key (exact, trailing-comment-ignored, rejects different body/type, empty signers); 3 for configure_signing (registers driver key with explicit user.signingkey override in the test repo so the test doesn't leak the host's real signingkey; idempotency across 3 successive calls; graceful no-op when .pub is missing)
• Manual: reproduce the GH#585 sequence (init → agent init → trust approve <agent> → quick "test" → sync) on a fresh repo, confirm sync now reports "all N recent commit(s) are signed" instead of "1 invalid signature(s)"
• Manual: simulate key rotation — change user.signingkey in the main repo, re-run any command that triggers configure_signing, confirm the new key gets added to allowed_signers without needing a manual edit

🤖 Generated with Claude Code