Row 85 round 2 (#1579), Filbert: **approve**.

Candidate manifest SHA-256: `bd7ea899c7cacef43b415b8e9ecb164e1f36718ffbe64347881a2c5653170c98`. Base: `79e126be0`. Packet: `agents/rocko/work/queue-85/r2/`. This answers request 28253 and follows my round 1 comment 28243.

I rebuilt the candidate in my own detached worktree at 79e126be0 and applied build.patch. All 11 manifest hashes match, `git diff --check` is clean, and none of the 11 paths changed between 79e126be0 and the current refactor head. No added line has an en or em dash, a curly quote, or a bidi or control character. To isolate round 2, I also applied the round 1 patch to the same base and diffed the trees. Round 2 changes only scripts/bus-service.sh, cli/tests/stack-copy.test.mjs, cli/tests/host.test.mjs and docs/TOOLS.md.

### B1: fixed

`install_guard` no longer calls `systemctl show` at all. `status` shows only the concrete `mosaic-bus@<business>.service`. The guard reads every directory from `systemd-analyze --user unit-paths` and globs `mosaic-bus@*.service.d/*.conf`; the `*` also matches the empty instance, so the template's `.d` is covered. A failed or empty path list, a relative path, or an unreadable directory, drop-in directory or file refuses before `mkdir`. Busy detection still uses list-units, which is right.

Live check (read-only): I used a stand-in `systemctl` that answers only `list-units`, with an inactive instance, and fails on any other call. With the real `systemd-analyze`, the candidate's `install --dir <scratch> --no-reload` refused with `drop-in /home/jwoltje/.config/systemd/user/mosaic-bus@mosaic-stack.service.d/pm.conf overrides ExecStart or WorkingDirectory`, and wrote nothing. Across all 21 unit paths on this host, that pm.conf is the only `mosaic-bus@` drop-in. So once Mos removes it in step 3, step 4's install will get past the guard.

The new tests cover both gaps. One uses a fake `systemctl` that fails `show` on the template the way systemd 261 does, asserts that no template `show` call is made, and asserts a template drop-in still refuses with the unit file unchanged. The other puts an unlisted instance drop-in in a second lookup directory and asserts that a failed `unit-paths` refuses. Rocko's template-show-regression and miss-unlisted-instance mutants are killed by those assertions, according to the packet receipts. I did not rerun mutants because Sage holds the slot.

### N1: fixed

TOOLS step 1 now saves the wrapper bytes and mode, the bus and Console units and drop-ins, the effective status output, the `--claude` pin and the Console flags. The first rollback restores that mixed setup in an order that works: it reinstalls the checkout bus without `--copy` or `--pm` while pm.conf is absent, then restores pm.conf. It states that the shared-copy check will refuse the restored setup and that #225 stays open. Every live step still needs Mos's go.

### N2: fixed

The proof fixture now runs the copy's real `startBroker` in a child process, so `runtime.mjs:68` derives `cliPath`. The child uses `mock.module` only to wrap `verifyHuman` and force `ppid` to 1 through its existing `readProcess` option; argv, pid, start time and nonce stay real. `runtime.mjs` and `human.mjs` are unchanged in the candidate. Rocko's runtime-wrong-proof-entry mutant points the runtime at a wrong CLI path and fails with `mosaic: human-required`, so the test now binds the runtime's own path. N4: TOOLS has no en dash left.

### Tests I ran (each file alone, no slot)

| File | Pass | Fail |
| --- | ---: | ---: |
| cli/tests/stack-copy.test.mjs | 9 | 0 |
| cli/tests/host.test.mjs | 21 | 0 |
| webui/tests/pinned-copy.test.mjs | 10 | 0 |
| webui/tests/service.test.mjs | 23 | 0 |

I set DOCKER_HOST to a missing socket and linked the canonical node_modules. No process was left in my scratch tree afterwards.

### Note (non-blocking)

- N1. The guard scans only `mosaic-bus@*.service.d`. systemd also applies prefix drop-ins (`mosaic-.service.d/`) and the all-services `service.d/`, and either could set WorkingDirectory. The brief names `mosaic-bus@` drop-ins, and neither directory exists on this host, so this is a note, not a blocker.

I ran no full suite, mutant or queue-commit.sh. The only live reads were `systemd-analyze --user unit-paths` and a directory listing. Sage's gate rerun is the check of record, and every live step still needs Mos's go.