**Test-only.** One file, 2 obsolete tests + 1 dead import removed. No
production change.
`project-engine-manager.test.ts` has been **red on main: 2 failed / 44
passed** → now **44 passed**.
## The failures
Both threw `TypeError: Cannot read properties of undefined (reading
'acquire')`, because both reach `(manager as any).globalSemaphore` — a
private field that no longer exists.
`project-engine-manager.ts:88` records why (`FNXC:CapacityModel
2026-07-28-20:10`, *"drop the cross-project cap"*):
> The shared cross-project semaphore, its mutable limit and the
`concurrency:changed` subscription are **DELETED**. Capacity is two
numbers per project; a machine-wide cap was a third limiter with its own
separate authority (a central-DB singleton row), and reconciling it
against the per-project gates is exactly the multi-limiter arbitration
this simplification removes.
So both tests assert residual-slot accounting on a shared pool that was
**deliberately** removed — not a regression.
## Why deleted rather than repaired
There is no shared semaphore left for them to describe. Reconstructing
one inside the test would assert a capacity model the engine no longer
has — a test that passes while describing fiction, which is worse than
the red it replaces.
Also drops the now-dead `ScopedAgentSemaphore` import (these were its
only uses). Lint does not flag unused imports here, so it would
otherwise have sat as quiet dead code.
## What I did NOT take, and why
`workflow-graph-optional-step-fix.test.ts` — the other red file adjacent
to this lane, 5 failures. Its failures are **U11 column-vocabulary
drift**: the replan rebound now resolves to `todo` where the test
expects `triage`, and one case gets a hard-cancel pause-abort log
instead of the Plan Review replan message.
That is the U11/U12 owner's semantics to settle. Picking whichever
column makes the assertion pass could silently encode the wrong
lifecycle target — and per the graph-entry contract doc, a rebound
landing in a column the workflow does not declare is precisely the
failure mode that "does not fail a test; it disables a recovery path in
production." Flagging it rather than guessing.
## Running tally of this cleanup thread
| File | Before | After |
|---|---|---|
| 27 logger mocks (#2573) | 206 failed | 4 failed |
| `merge-error-recovery` (#2559) | 10 failed | 0 |
| `reviewer` (#2547) | 2 failed | 0 |
| `project-engine-manager` (this) | 2 failed | 0 |
Every one was a test describing behavior that had moved or been deleted,
or a mock that had drifted from its real shape — none was a product
defect. That pattern is worth naming: on this repo a red non-blocking
suite has mostly meant *stale tests*, which is exactly what makes it
easy to ignore, and exactly why it silently corrupted my own safeguard
measurements in #2511.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
c92bce2f8c