backlog: record #1103's second occurrence, on Linux, and retire the Windows framing

[wshallwshall]

Amends BACKLOG #1103, filed hours ago in #290. Ledger-only.

Why this matters more than a second data point

#1103 was found on Windows, where the failure reports WinError 10013 — access forbidden. PR #287, a CodeQL action SHA bump touching two workflow files and no Python, failed the same way on ubuntu-latest:

ConnScaleError: engine exited during startup:
  [Errno 98] bind on ('127.0.0.1', 45382): address already in use
1 failed, 10453 passed, 1147 skipped in 849.41s (0:14:09)

Port 45382 is Linux-ephemeral and, like 62748, sits far outside the inbound window [20000, 30000) — the API family again. The step ran 14:19 against ubuntu's 25:00 cap, so not a timeout there either.

The defect is platform-independent; only the errno differs. Linux reports EADDRINUSE — "address already in use" — which names the cause. Windows reports 10013 — "access forbidden" — which does not. The item was filed under the misleading spelling. Both are now recorded, because the same defect presents as two unrelated-looking failures, and a reader who meets only the Windows form reaches for permissions and runner images rather than port allocation.

Left unrecorded, the natural next move is to file the Linux failure as a separate item — which is the conflation ci.yml's own scope note warns about, and which this session has already had to untangle four times on one leg.

Both occurrences are on PRs that cannot reach the code

PR	content	leg	error
#289	CI timeout values	windows-2022	10013
#287	CodeQL SHA bump	ubuntu-latest	98

That is the signature of a harness defect rather than a regression, and it is why neither was re-run before being diagnosed.

Verification

Banner invariant holds. tests/test_backlog_status_check.py: 19 passed. Repo-wide link gate: 5,359 links / 347 files, all resolve.