fix: two PowerShell unroll defects -- install-gate -Status crash, and a phantom overlap occupant

[wshallwshall]

Two unrelated instances of one PowerShell defect, in separate commits so either can be dropped alone.

1. install-gate.ps1 -Status crashed (the brief)

Get-WiredMatchers built a HashSet and ended with a bare return. PowerShell unrolls an enumerable return value into the pipeline, so no caller ever received a set:

matchers	actually returned	consequence
0	$null	.Contains() at the wire-set scan throws; -Status exits 1
1	[String]	.Contains() becomes a SUBSTRING test, silently
2 or more	[Object[]]	membership, which is why this went unnoticed

The one-element arm is why this is a fix and not a crash report. Measured against the shipped rule set, with a config dir wiring exactly one matcher NotebookEdit:

OLD  exit 0   UNWIRED: Agent, Bash, MultiEdit, PowerShell, Task, Workflow, Write
NEW  exit 0   UNWIRED: Agent, Bash, Edit, MultiEdit, PowerShell, Task, Workflow, Write

Edit is a strict substring of NotebookEdit, so it was swallowed. Exit code 0 both times: a clean, confident wrong answer from the audit whose only job is to report which rules never fire.

The empty arm also broke the independent audit below it, which tests $null to mean "Read-Settings threw" and prints settings.json is not valid JSON. Zero matchers returned $null too, so a readable, simply-unwired directory was reported as unparseable:

OLD  UNREADABLE: .claude-backup  <- settings.json is not valid JSON
NEW  not judged: .claude-backup (not a launcher name) -- carries no gate wiring

That branch is correct again as a consequence of the fix, and now says so in a comment.

Not a CI regression. A hosted runner has no ~/.claude, so $ConfigDir resolves empty and the loop never runs. It fires on developer machines, hardest on ones carrying a second .claude-account-* directory.

2. overlap.ps1 printed a phantom occupant (found while sweeping)

Same class, but it survives to disk. Build-Map returns an empty array, PowerShell unrolls it to AutomationNull, and in process that is harmless. It does not survive serialisation: ConvertTo-Json writes "rows": null, the next run reads it back with @($c.rows), and that is a one-element array holding $null. Count is 1, the zero-rows guard does not fire, and the report prints a blank worktree on a blank branch with 0 changed files.

collision_gate.ps1 runs overlap.ps1 -Json on every gated edit and the cache is written before the -File early exit, so the hot path arms it constantly and any bare overlap.ps1 inside the cache window inherits it. An invented collision is not a safe direction to fail in: it is the answer people work around, and a report that cries wolf on a quiet repo is one they stop reading.

Tests

Four new tests, three for the install-gate arms and one for the cache round trip. Each was verified RED on the unfixed script for its own reason rather than incidentally: the empty arm by the null-method crash, the substring arm by content on an exit-0 run, the audit arm by the false UNREADABLE line.

The install-gate tests plant a copy of worktree_gate.ps1 in the synthetic home. Without it $GateDst does not resolve, the implemented-rule set is empty, and every assertion about UNWIRED passes vacuously. Both runs report against 10 implemented rule(s), which is what makes the fixture non-vacuous.

Red/green was proved against a scratch copy of the script rather than by stashing the live file, so no concurrent reader could observe a tree with the fix missing.

63 tests pass across the gate, worktree and coord files. ruff format, ruff check and mypy clean.

Generated with Claude Code