The fixture board, a host client, and three bugs that shipped in 0.1.0

[Boernsman]

Works through the tasklist, and fixes three bugs found along the way that
were all present in 0.1.0.

Large, so the three commits are meant to be read in order. The first is a
data fix to the example board, the second is .gitignore, and the third is
everything else.

Three bugs that shipped in 0.1.0

pinside init did not work at all. The positional board and the
carrier option --board shared an argparse destination, so every invocation
looked the .kicad_pcb path up in the module catalogue and refused. CI's
end-to-end job runs that exact command. The option is now --carrier, and
main() derives its command list from the parser rather than a hand-kept
set, which is what let a newly added subcommand be swallowed too.

pinside project --board pico2w emitted a schematic KiCad would not
open. KiCad writes the Pico W as a symbol deriving from the Pico, carrying
properties and no pins; copied into a schematic's lib_symbols that
extends resolves to nothing and every net in the design is isolated.
Derived symbols are flattened against their parent now. The units have to be
renamed too, and missing that gives you Failed to load schematic and no
further explanation, which took a file-level bisect to find.

The example board's netlist could not survive being opened. All sixteen
test points were written with net ordinal 1 under sixteen different names.
Through KiCad 9 a net is identified by its ordinal and the name is a
label, so KiCad read them as one net and dropped fifteen on the first save.
pinside check called it clean, because the reader takes the name off each
pad and never compares ordinals. That is now PS043.

A board with a slot was reported as having an unclosed outline
(PS002, an error) on geometry any fab would cut. Edge.Cuts holds every
edge a board has; the largest ring is the perimeter now and the rest are
cutouts.

New

• pinside.client and pinside probe, an optional extra
  (pip install 'pinside[client]'). Refuses a fixture whose config hash
  disagrees with the config it was handed, and classifies responses the way
  JSON-RPC 2.0 defines them rather than by id: an echoing port otherwise
  looks like an answer. Found by pointing the client at pyserial's own
  loop://.
• Baselines. --write-baseline records the findings a board has been
  judged on; --baseline accepts them. Suppression is by code and by
  reference, so a new occurrence of an accepted code still fails, which is
  what makes the file safe to commit.
• --json on generate and project, including the refusal path.
• Eight checks. PS003, PS004 (cutouts, second outlines), PS013,
  PS014 (probes and holes over a cutout), PS027 (receptacle body fouling
  a component), PS033, PS034 (rails and reset lines the fixture cannot
  reach), PS043, PS044 (net numbering KiCad will not preserve).
• A GND pour on the generated fixture board, and a generated README that
  turns the plate force into hardware: clamp or thumbscrew, load per
  standoff, and how long the standoffs must be to leave the probes travel.
• rp2354a, probe provenance, PyPI trusted publishing, coverage at 87%,
  SECURITY.md, issue and PR templates.

How the net checks were decided

PS043 and PS044 claim KiCad loses nets, so I asked KiCad rather than
assuming. kicad-cli pcb export ipcd356 is its own answer to what a board's
netlist is. Five malformed netlists went through it:

Case	KiCad's verdict	Check
One ordinal, several names	keeps the first, rest become N/C	PS043 error
Named net on ordinal 0	N/C; net 0 is the no-connection net	PS044 error
Net table contradicting a pad	that pad becomes N/C	PS043 covers it
One name, several ordinals	merges them into one net	none, agrees with pinside
Pads on undeclared ordinals	reads back correctly	none, not a defect

Three of the five turned out not to be defects and got no check.
tests/test_kicad.py runs the same boards through ipcd356 and compares,
including the merge case pinside deliberately does not flag, so that
decision is confirmed rather than assumed.

Testing

208 → 221 tests, and the two CI-invisible paths are now covered: the KiCad
tests run inside the kicad/kicad:10.0 container with
PINSIDE_REQUIRE_KICAD=1, which turns a skip into a failure so the job
cannot go green having run nothing, and the generated firmware is
cross-compiled against a pinned Pico SDK.