Two CLAUDE.md style violations to fix before merge.

1. CHANGELOG.md — four sentences packed onto one continuation line

CLAUDE.md is explicit: "CHANGELOG.md bullets — one sentence per line within a bullet."

The new fix(kernel) entry's continuation line contains four sentences on a single line.
Split into one sentence per line:

- **fix(kernel): persist per-agent skill & MCP-server assignments to `agent.toml` so they survive a restart** (#6045) (@DaBlitzStein).
  `set_agent_skills` and `set_agent_mcp_servers` wrote only the SQLite blob and never called `persist_manifest_to_disk`, unlike the sibling setters (`set_agent_model` / `set_agent_channels` / `set_agent_schedule` / `set_agent_tool_filters`).
  Since boot reconciliation re-syncs each agent from its on-disk `agent.toml` and overwrites DB-only fields, a skill/MCP assigned via the dashboard returned `200`, showed as assigned in the API, then was silently wiped on the next daemon restart.
  Both setters now persist to disk.
  Regression test: `test_set_agent_skills_persists_allowlist_to_agent_toml`.

2. crates/librefang-kernel/src/kernel/agent_state.rs — two multi-line // comment blocks

CLAUDE.md: "Never write multi-paragraph docstrings or multi-line comment blocks — one short line max."

The first block (8 lines before set_agent_skills's persist_manifest_to_disk call) and the second (4 lines before the MCP one) both violate this.
The WHY is non-obvious and worth a comment, but it should fit in one line each.

Suggested replacements:

// Without this, boot reconciliation silently overwrites DB-only fields from the on-disk agent.toml on restart.
self.persist_manifest_to_disk(agent_id);

// Same persistence requirement as set_agent_skills: boot reconciliation overwrites DB-only fields from agent.toml.
self.persist_manifest_to_disk(agent_id);

The detailed mechanism (boot.rs log message, sibling list) belongs in the PR body — where it already appears — rather than in an inline comment that will rot as the codebase evolves.

Generated by Claude Code

One finding during daily review pass:

Missing regression test for set_agent_mcp_servers

The PR adds persist_manifest_to_disk to both set_agent_skills (lines ~460–475 of agent_state.rs) and set_agent_mcp_servers (lines ~534–540), and correctly adds test_set_agent_skills_persists_allowlist_to_agent_toml for the skills path. But there is no parallel test for the MCP-servers path — a future regression in that setter could go undetected.

Please add test_set_agent_mcp_servers_persists_servers_to_agent_toml alongside the existing test, using the same pattern (spawn agent, call set_agent_mcp_servers, read_to_string the agent.toml, assert the server name appears). The two paths got identical fixes; they should have identical coverage.

No other issues found.

Requesting changes to consolidate the three unresolved review points so this doesn't merge with known gaps. All are small and the fix logic itself is correct — just needs the coverage and style cleanup before merge.

1. Missing regression test for set_agent_mcp_servers (raised 06-10 and 06-12, still open). The PR fixes both set_agent_skills and set_agent_mcp_servers, but only the skills path has a test. Add test_set_agent_mcp_servers_persists_allowlist_to_agent_toml mirroring the existing one: boot kernel, spawn agent with a known source_toml_path, call set_agent_mcp_servers([<server>]), read the on-disk agent.toml, assert the server name is present. Two identical fixes should have identical coverage.

2. CHANGELOG continuation packs four sentences on one line (raised 06-11). CLAUDE.md: "CHANGELOG.md bullets — one sentence per line within a bullet." Split the new fix(kernel) entry's continuation into one sentence per line.

3. Two multi-line // comment blocks in agent_state.rs (raised 06-11). The 8-line block before set_agent_skills's persist_manifest_to_disk call and the 4-line block before the MCP one violate the kernel crate's one-line-per-comment convention. Keep the WHY but compress each to a single line; the detailed mechanism (boot.rs log, sibling list) already lives in the PR body.

None of these are logic blockers — the persist fix is correct and the breadth is complete (skills + mcp were the only two setters missing the call). Re-request review once they're addressed.