**Self-reported regression from #2772, which has already merged.** The
seam gate on main currently prints a clean bill of health while five
real omissions sit in front of it.
## What broke
Closing the imported-shadow hole in #2772 taught the check to record
which module each callee was imported from, and to exclude a call site
whose module basename does not match the seam's declaring module. That
works for relative imports. It does not work for barrel imports.
Engine and CLI reach core through `import { ... } from "@fusion/core"`.
That specifier's basename is `core`, which never matches a module name
like `near-duplicate-canonical` — so **every barrel-imported call site
was classified as "a different function of the same name" and dropped.**
The check stopped seeing engine's and cli's calls into core entirely,
which is most of the cross-package surface it exists to watch.
## Measured
On `main` today:
```
[check-inert-flag-seams] 21 lane/flag seams, all supplied at every production call site.
```
With this fix:
```
packages/core/src/near-duplicate-canonical.ts: isNearDuplicateCanonicalInactive()
— supplied by 6/11 call sites; omitted at
packages/engine/src/self-healing.ts (x2), packages/engine/src/triage.ts (x3)
```
Those five were always there. Earlier I reported this seam as "supplied
by 5/6" — that number was wrong for this reason, and the engine sites
were invisible to me when I said it.
## The rule now
Only a **relative** specifier identifies a module well enough to exclude
a call site on. Anything else is unresolved, and unresolved must mean
**counted**: an over-counted seam produces a false report somebody
investigates, an under-counted one produces silence. I had this
backwards, and it is the second time in this lane a change made the gate
read cleaner while catching less.
## Both directions verified
- **Barrel imports counted** — the five engine sites appear.
- **Relative shadows still excluded** — lifting the
`sortTasksForDisplayColumn` entry still reports it unsupplied, so
`Lane`/`Board`/`ListView` calling the dashboard twin through
`"./taskSorting"` does not clear core's seam. That was the entire point
of the original fix and it still holds.
## The five engine omissions
Not fixed here — engine-owned, reported on #2785. Same shape as the
merge-queue bug in #2819: a canonical resting in a **renamed active
column** reads as *inactive*, so duplicate markers get cleared against
live work.
They carry TEMPORARY per-file exemptions so this PR is green and
self-announcing. Noted at the entry: the key is `<file>::<function>`, so
a file with two omitting calls is exempted for both — coarser than I
want, recorded rather than left to be discovered.
## Verification
`pnpm test:gate` green, lint 0.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
fe8f3af751