Defer the box-tree imports in cli/impl so the modules import from the wheel

[danielrmerskine]

Closes #241.

What was wrong

cli/impl/box_config.py:13 and cli/impl/power/enable_disable.py:6 imported lager.* at
module scope. That package lives under box/ and is not shipped in the lager-cli wheel, so
both modules raised ModuleNotFoundError on any clean pip install. They worked only where
box/ happened to be on sys.path — dev checkouts via the test PYTHONPATH convention, and
boxes.

Why this is a small fix and not a code move

Worth stating up front, because the issue offered "move the needed code under cli/" as an
option and that would have been a mistake:

• These files are uploaded to the box and executed there (cli/core/net_helpers.py:683-698,
  cli/commands/box/config.py:84-102). The CLI never imports them — the coupling is a filename
  string passed to get_impl_path(), not an import.
• They must keep shipping in the wheel, because get_impl_path() resolves them from the
  installed package on disk. Excluding them would break lager logic and lager box-config on
  a pip install.
• So the only thing that was actually wrong is when the import happens. Moving Net under
  cli/ would have dragged 100 modules, ~28.6k lines and 18 hardware SDKs into a CLI wheel
  whose install_requires is 16 pure-Python packages.

The fix is the pattern already in the tree two directories over: cli/impl/measurement/scope.py
imports Net/NetType inside get_rigol_net() and get_source_net() rather than at module
scope. Both defective files now do the same. box_config.py's sys.path.insert('/app/lager')
moved inside __main__ alongside its import, since it is meaningless off-box.

The import path is deliberately unchanged (from lager.nets.net import Net, NetType, not
from lager import ...), so on-box resolution is byte-for-byte what it was. Only the timing
moved.

The regression test matters more than the fix

The fix itself is six lines moved. What makes it stick is test_impl_host_importable.py, which
imports every cli/impl module — discovered from disk, not hardcoded — in a subprocess with
box/ stripped from sys.path and a meta_path finder that raises on any lager import.

That subprocess isolation is the point. The unit suites run with box/ on PYTHONPATH
(unit-tests.yml), which is exactly what hid this defect from every existing test. A same-process
test would have passed regardless of whether the bug was present.

Fault-injected to confirm it is not tautological: restoring the module-level import makes it fail
with cli.impl.power.enable_disable does not import without the box tree, and the failure message
points at the scope.py precedent. Restoring the fix makes it pass.

Also in here

Two comments in box/lager/http_handlers/usb_scanner.py instructed the reader to mirror changes
into cli/impl/query_instruments.py, which was deleted in the :9000 migration. Following them was
impossible. One is dropped, the other now records that this is the only remaining copy. The
neighbouring "Extracted from…" / "previously the … script executed over :5000" comments are
untouched — those are accurate history rather than stale instructions.

Verified

• Packaging gate green end to end: build → twine check (both artifacts PASSED) → wheel into a
  fresh venv → import walk. 177 modules walked, 31 failures, 0 unexpected. All 31 are the
  vendored-elftools entries that Remove the vendored elftools tree and the unreachable GDB source-archive path #249 removes; the two cli.impl failures are gone.
• The baseline is two-sided, so deleting those two lines is itself the proof: had either module
  still failed, or had the entries been left behind after the fix, the gate would have failed.
• unit (cli): 1232 passed, 2 xfailed (was 1224 + 2; +8 from the new test).
• tools/check_coverage_counts.py passes, including the per-file inventory row. unit (box)
  reports as platform-gated here (7 tests need /proc and flock(1); this is macOS) — untouched
  by this change and checked on Linux CI.
• ruff --select E9,F63,F7,F82 clean.

Note for whoever merges second

This and #249 both edit tools/packaging_import_baseline.txt, and this, #249 and #251 all add a
## [Unreleased] CHANGELOG section. The baseline hunks are disjoint (elftools block vs cli.impl
block) and should auto-merge; the CHANGELOG sections will want a trivial manual merge into one.
Once both this and #249 land, the baseline holds no functional entries at all — only the
cli.setup note explaining why that one is deliberately unlisted.