muse commit --sign is agent-only, silently no-ops for human commits

Background

Discovered while recording Build with Muse episode 04, "Identity, Without Passwords" — the whole point of the episode is that any identity (human or agent) can sign a commit with its Ed25519 key, no password required.

Repro:

muse auth whoami --json
# {"type": "human", "handle": "gabriel", "key_set": true, ...}

muse commit -m "adding README" --sign
# [main sha256:91e845a8...] adding README
# 1 file changed (1 added)
# -- no warning, no error --

muse read-commit sha256:91e845a8... --json
# "signature": "", "signer_public_key": "", "signer_key_id": ""

muse verify-commit sha256:91e845a8... --json
# "valid": false, "signer": "", "key_status": "unknown"

A real, registered human identity exists (confirmed via muse auth whoami and muse auth show), yet --sign produced an unsigned commit with no indication anything went wrong.

Root cause

muse/cli/commands/commit.py gates the entire signing-resolution block on both sign and resolved_agent_id:

if sign and resolved_agent_id:
    from muse.cli.config import get_signing_identity
    signing = get_signing_identity(root, agent_id=resolved_agent_id)
    ...

A human commit has no --agent-id/MUSE_AGENT_ID, so resolved_agent_id is "" (falsy) — the whole block is skipped. signing stays None, pre_signer_public_key stays "", and the commit is written unsigned.

This is unnecessary: get_signing_identity() already supports human-key resolution when agent_id is None/falsy (see its docstring, resolution order step 3: "Human entry in ~/.muse/identity.toml keyed by bare hostname"), and resolve_signing_identity() already treats falsy agent_id as "skip the agent-specific lookup, go straight to the human entry." sign_commit_record() also accepts an empty-string agent_id fine — it's metadata-only, not part of key resolution. The gate at the call site is the only thing preventing human signing from working.

Fix

Change the gate to if sign: and pass agent_id=resolved_agent_id or None so an empty string and None are treated identically all the way down (defensive, even though both already resolve the same way today).

Acceptance criteria

• muse commit -m "..." --sign as a human with a registered identity produces a commit with non-empty signature/signer_public_key.
• muse verify-commit on that commit reports "valid": true.
• Existing agent-signing behavior (--sign --agent-id ...) is unchanged.
• See companion issue #2 for the silent-failure behavior when no identity is configured at all.