PerfStudio v0.10.0

A polish release: nothing new to learn, and a long list of things that were almost right.
Four audits — the window, the three views, the engine's own prose and the loader's
tolerance for bad files — each came back with what they found, and what follows is what
survived being checked.

Added

• The three panels that had no way back. Parts, Nets and DRC / LVS can be closed from
  their title bars, and the only route back was a right-click on the menu bar, which nobody
  finds. View ▸ Show Parts (Ctrl+1), Show Nets (Ctrl+2) and Show DRC / LVS (Ctrl+6)
  join the three toggles the other panels already had.
• Right-clicking a conductor. The menu over a trace was the bare-board menu — Paste,
  Connect, Fit — with no way to delete the trace under the pointer. It now selects the
  conductor and offers Copy, Duplicate and Delete, the way a right-click on a part already
  selects the part.
• Ctrl+0, Ctrl++ and Ctrl+- follow the keyboard focus. Over the schematic panel
  they fit and zoom the sheet, which had a Fit button and no key; over the board they do
  what they always did. The sheet's zoom also gained the board's limits — a few trackpad
  flicks used to leave a blank field or one pin, with nothing but the button to recover.
• Undo and Redo say what they will do. The Edit menu read "Undo" and "Redo" and the
  answer was one hover away; it now reads "Undo Move R4 to C7", and Redo — which never
  named anything, because the bus had no way to ask — reads the same. CommandBus gained
  redo_history() for it.
• Load warnings are shown, not counted. Opening a file that was read differently from
  how it was written — a diagonal solder-trace step, an unknown property, a duplicated
  reference — put "(2 warning(s))" in the status bar and nothing else anywhere. They are
  listed in a dialog now. And a duplicated reference IS a warning now: every net node
  names a part by reference, so two R1s is a netlist that cannot say which one it wired.
• Two DRC rules for the file only a hand can write. conductor-off-board (a conductor
  with a hole outside the grid, the twin of component-off-board) and unknown-footprint
  (a part whose footprint neither the library nor the id grammar can produce). No command
  can create either state, so both only ever fire on a document edited by hand or written
  by another tool — which is exactly the file that must still open and be told what is
  wrong with it. Both are Python-only (PYTHON_ONLY_RULES); the second turned out to fire
  on eight of the fifteen golden fixtures, all on one id, c-disc-1, which exists in
  neither engine. The fixtures are dumps of the original and are left as they are; the
  rule is excluded from the differential comparison and pinned by its own test.
• The interface is Turkish where it was still English. The status bar's DRC, LVS and
  ratsnest fields, the Nets panel's "done", the DRC panel's summaries, the schematic
  panel's summary, the three file dialogs, the progress dialog's five labels, the
  confirmations' bodies, the "select a part first" line and its five verbs, the Board Setup
  material list, and the board's own prose while measuring, picking pins and joining them.
  tests/test_i18n.py now also checks the View menu for accelerator clashes, which found
  one in each language.

Changed

• A group moved together is one undo step. Five parts dragged or nudged at once used
  to be five component.move commands, so five Ctrl+Z presses to put them back — and a
  refusal part-way through left the group torn apart. The scene dispatches one
  component.moveMany, which is all-or-nothing: a locked part in the selection refuses the
  whole move and says which part.
• Every destructive question puts Cancel under Enter. Delete parts, delete conductors,
  delete a net and reload from disk all used QMessageBox.question with no buttons named,
  which puts Yes under Enter — so a confirmation answered by reflex was a deletion. They
  share one box now, MainWindow._confirm, whose button carries the verb ("Delete") and
  whose default is Cancel. Tests that used to stub QMessageBox.question stub _confirm.
• Deleting a mixed selection deletes all of it. A rubber band that took a part and the
  two traces on it used to delete the part and keep the traces without a word; the
  question now says the conductors go too, and they do.
• The loader refuses what board.set refuses. A .perf with "pitch": 0, no columns,
  or a drill wider than its pad used to load without a word, and then every repaint divided
  by it — in the pure render path, a segfault. Two components sharing an id (a part no edit
  could ever reach, and one the placer refused the whole board over) and a formatVersion
  below 1 (no such version has existed) are refused too, with the same codes and paths the
  other load errors carry. Missing required field "cols" now says "board.cols", the
  path every neighbouring message already quoted.
• CommandBus.dispatch never raises, as its contract says. A payload of the wrong
  shape — a dict where a dataclass was expected, a tuple where a HoleCoord was — used to
  surface as an AttributeError out of the command; it is {ok: false, code: "invalid-payload"} now, which is what an agent on the far side of the MCP server needs.
  The undo stack is bounded at a thousand entries, where it had no bound at all.
• Save is enabled only when there is something to save. Ctrl+S on an unmodified
  board rewrote the file — a new modified stamp, a new mtime — for nothing. An untitled
  board can always be saved somewhere, so Save stays live there.
• The Downloaded strip has a Close button. Its only button was Hide, which skips that
  version for good — so closing the strip after a successful download told the application
  never to mention the release the user had just fetched. Hide stays on the announcement,
  where it means what it says.
• board.set names every stranded part, and refuses to strand a track cut. Shrinking a
  board with three parts outside the new size used to name the first, so the refusal
  repeated once per part, each discovered only after the previous one had been moved. A
  cut off the board was kept silently — and if the board was grown again, it broke a strip
  nobody remembered cutting. A component also needs a non-empty reference now; the wiring
  cannot name one without.
• MCP results mean one thing by ok. run_lvs returned ok: false when the board
  did not match the schematic, while everywhere else ok: false means the call was
  refused; it returns ok: true with matches_schematic now, and run_drc carries
  ok: true too. autoroute on stripboard returned cuts: 0 for an empty plan and
  cuts: [...] for a committed one; it is always a list, with cut_count and
  link_count beside it. export_pdf with no directory wrote two PDFs into the
  server's working directory, against the instructions' own promise that nothing writes to
  disk unless a path is named — it refuses with no-directory. Paths accept ~, and an
  empty path is refused rather than blamed on permissions. The server's instructions now
  list all six connection kinds (not four), give the design-first working order
  docs/MCP.md gives, say what plain is as a board type, and send an agent to reroute
  after optimize_placement rather than to autoroute, which only adds.
• The KiCad importer says what it was handed. A .kicad_sch given to it — the
  schematic instead of the netlist exported from it, which is the first-time mistake — got
  "no export form found"; it now says it is a schematic and where the netlist comes from.
  Two nets sharing a code (a broken export, but the user's export) used to refuse the
  whole import as a duplicate id; they are imported with a suffix and a warning. A
  component with no footprint is warned about, since it is what forces the host to guess
  one.

Fixed

• Measuring, then flipping the board, broke every tool. The measurement marker is an
  item like the others and a rebuild destroys it, but the mode kept a wrapper for it — so
  the next click, or the next arming of any tool (all of which disarm measuring first),
  raised from removeItem and left every board mode unreachable until the window was
  reopened. test_measuring_survives_a_flip is the reproduction.
• Three more things a rebuild lost. A half-drawn trace kept its path and lost its
  picture, so after a flip or an undo the tool was armed with nothing on the board and
  Enter committed an invisible chain. The placement ghost came back at A1 after every
  placement — at exactly the moment the next part was being lined up. And a conductor
  selected to be deleted was deselected by whatever unrelated command came first, while
  parts had always been re-selected by id.
• Picking a part mid-trace armed both. Clicking the Parts panel while a trace was being
  drawn left the ghost following the pointer while every click still extended the trace,
  so the part could never be placed. The board modes are mutually exclusive and placement
  now disarms drawing, as drawing already disarmed placement.
• The solder side and the keyboard disagreed. The arrow keys moved a part one column in
  the document, which on the mirrored solder side is a step to the LEFT under Right;
  Rotate Clockwise turned the other way there for the same reason; and the placement ghost
  painted a DIP's pins running one way while the placement put them the other. A drag was
  right all along, because it round-trips through screen_to_hole. All three now correct
  for the side, and test_arrow_keys_follow_the_screen_on_the_solder_side holds it.
• Flipping the board flips the view and the camera with it. The scene is mirrored
  about the hole span's midpoint, so a flip while zoomed in on the left edge showed the
  right edge; the viewport centre is mirrored too now. And the 3D panel, whose own
  docstring said the camera turns over "when the board is flipped", stayed on the component
  side — it turns over.
• The last build step was unreachable on the assembly slider. With N steps the slider's
  maximum was N, and the maximum means "the finished board", so step N had no position of
  its own: picking it in the Build Guide panel showed the finished board with nothing picked
  out. The range is N+1.
• Delete was greyed out for a conductor on its own. Which is the one selection
  conductors were made selectable for. The status bar also said nothing about a
  conductor-only selection while Copy and Duplicate were live for it.
• Save could not fail. Ctrl+S into a read-only folder or onto a full disk was an
  unhandled traceback with the board still unsaved behind it, and the close guard would
  have treated it as saved. It is a dialog, the close guard is told, and a failed Save As
  does not adopt the path it failed to write.
• Exports that reported success for files that were never written. The 1:1 PDF export
  trusted QPdfWriter, which reports an unwritable target as a warning on stderr and
  returns; the 3D snapshot trusted vtkPNGWriter, which does the same — and the snapshot
  had none of the "is there GL here" guard the guide export has, so on a machine without
  offscreen GL it ended the process with every unsaved edit in it. Both check the file on
  disk; the PDF export offers to open its output like the other two exports.
• A freshly opened board inherited the previous one's "parts moved" marks. Net ids are
  document-local — every board has a net-1 — and the record of which nets were routed for
  a position a part had since left was never cleared on Open, so Ctrl+R on a new file
  asked about re-routing nets nobody had touched. Open, New and Recover all forget it now,
  with the schematic selection, the described custom parts and the placement seed.
• Board Features ▸ Remove with nothing chosen deleted the first row. Chosen for the
  user. The button follows the selection now.
• The Cut Track tooltip stuck on "only stripboard". A board switched to stripboard
  under an open window kept the explanation of why the tool was unavailable on a tool that
  had started working. Reset 3D Camera and Exploded View, which do nothing until the 3D
  panel has been opened, are greyed out until then rather than silently inert.
• Escape wiped the status bar when there was nothing to leave. The routing summary is
  posted with no timeout so it can be read, and a reflex Escape cleared it. It clears only
  when a mode was actually left.
• "Saved …" and "Exported …" never went away. Both were posted with no timeout, so the
  path of the last save sat in the status bar for the rest of the session. Eight seconds,
  like every comparable message.
• Menu tooltips were switched on for four menus of twelve. The explanations written for
  Reload, Board Setup, Board Features, the exports, Measure and the 3D items, and the full
  path behind each recent file, never appeared anywhere. Every menu shows them.
• The recent-files menu went stale in place. It was rebuilt on save, open and clear
  only, so a file moved or deleted while the window was up stayed listed and failed when
  picked. It is rebuilt as it opens.
• The recovery offer and the first-run update question stacked. Both are scheduled a
  moment after the window appears, and the second opened its modal on top of the first's.
  The update question waits while another dialog is up.
• A recovery record that would not parse was offered forever. And stood in front of
  every other record behind it. It is dropped after the failure to open it is reported.
• The "cannot write the recovery file" warning was wiped within seconds. It was a status
  message, and the next command's description replaced it — so autosave was dead for the
  rest of the session with nothing on screen saying so. It is a permanent field now, shown
  while the condition lasts and cleared when a write succeeds again.
• A file change under a running planner swapped the board out from under the plan.
  The planner pumps the event loop so the window can repaint and offer Cancel, which also
  let the file watcher reload the document mid-plan — and the plan was then committed into
  a different board. The watcher waits for the planner, and so does the close button, whose
  caller would otherwise have dispatched into a window that had already cleared its
  recovery record.
• A refused net dialog threw away what was typed. A duplicate name was refused into
  the status bar after the dialog had closed; it is a dialog now, and the form comes back
  filled in.
• The placement cursor was the only mode cursor. Drawing, connecting, cutting and
  measuring all left the ordinary arrow over a board where a click no longer selected
  anything; every mode shows the crosshair.
• Leaving the board cleared nothing. The status bar kept naming the hole the pointer
  left the board over, and Paste from the menu still landed there. Panning with the middle
  button and zooming from the keyboard also left the ghost and the address on the hole
  that had scrolled away; both re-read what is under a pointer that has not moved.
• Zoom stopped one notch short of its own limit. The last wheel step before a bound did
  nothing rather than landing on the bound, and a fitInView that landed outside the range
  — two pads framed from a DRC finding in a large viewport — left the wheel dead in both
  directions. Steps are clamped; fits are pulled back inside.
• Opening a file from the command line did not fit it. The dialog route fitted the
  board; perfstudio big-board.perf arrived at a fixed six pixels per millimetre, which is
  a corner of anything bigger than a 7 × 9 cm board.
• A truncated .net was a traceback. SExprSyntaxError was not a ValueError, which is
  what both importers catch, so the commonest real breakage escaped the window and the MCP
  server alike. It is a ValueError now.
• Double-clicking a pin while wiring the sheet opened a dialog over the half-made
  connection. The sheet follows the board's rule: a mode owns the first click of any pair.
• A download that could not be written was an exception in the event loop. A disk
  that fills part-way through 300 MB now cancels the transfer, reports why, and removes the
  .part file, like every other failure in the updater.
• The Schematic panel's Remove was enabled for a symbol nothing defines — one that
  exists only because a net names it — with an explanation of why it could not work as its
  only outcome. The empty panel also said "0 part(s), 0 net(s)" and nothing else; it now
  says what the two buttons under it are for.
• Tooltips learned the two things people hover for. A conductor's tooltip names its
  net, and a part's names its hole and rotation.
• The guide said several things that were not so. "Cut it to 32 × 22 holes (W × H mm)"
  computed the size from the pitch alone and dropped the printed border, so on a board from
  a preset it was 4 to 9 mm too small. The DIP step said to leave the IC "until phase 8",
  where there is no such step; it now says the socket goes in now and the IC after the
  closing checks. "Put U1 into their sockets" agrees in number and allows for no socket.
  Bare wire was told to be stripped 0 mm. "Flux is not optional on a run this long" was
  appended to two-pad joints. A trapped jumper was named cond-7, which nobody at a
  bench can look up; it is named by its ends. The polarity note's dash swallowed the hole
  address. The header spelled the board 32×22 FR4 twelve lines above 32 × 22 and
  FR-4. One separator for a part's span (→), one spelling for a wire's colour across
  the CSV, the HTML and the JSON (colour), one wording for its type. The goldens were
  re-blessed after reading every changed line.
• DRC, LVS and the router name things by address. Four DRC rules identified a run only
  by its conductor id; they carry its ends now, and the material is spelled FR-2 rather
  than the enum's FR2. mOhm is mΩ, as the guide already had it. An LVS short could
  print an internal physical-net id where an address was promised. Router explanations
  are capitalised consistently. The autoroute summary the headless CLI prints used a
  non-ASCII separator four hundred lines below the comment forbidding one; --headless
  also pluralises one way throughout, uses the window's own reader (so a directory is one
  line and not a traceback), and says so when it cannot create its output folder.
• The documentation caught up with the code. Both READMEs claimed v0.4.0 and
  forty-four MCP tools (it is v0.9.0 and fifty-one), recommended the python -m
  invocation docs/MCP.md warns against, and — in Turkish — gave English menu paths.
  CONTRIBUTING.md said ruff was not a gate and counted ~1260 tests; it is, and there are
  ~2070. examples/README.md said the FR-2 guide "halves the dwell"; it cuts it from
  three seconds to two.
• The AppImage build fetches its own runtime. v0.9.0's release job died on
  Failed to download runtime: server returned status code 302 twenty minutes after
  the same job had passed on a dry run: an AppImage is a squashfs image behind a small
  runtime, appimagetool downloads that runtime itself at build time, and its fetcher
  does not follow a redirect. curl -L does, and --runtime-file is the way round that
  appimagetool's own error message names.
  • It also closes a hole the script was already arguing against. appimagetool is
    pinned to a release because "whatever was on the server that morning" is not
    something a release can be reproduced from — and the runtime, which is the part that
    ends up INSIDE the artefact, was being pulled from continuous by something nobody
    could see. It is a visible input now.

---

Windows - PerfStudio_..._Setup.exe. It is unsigned, so SmartScreen will
warn on first run: More info -> Run anyway. A signed build is planned.

Linux - perfstudio-...-x86_64.AppImage. chmod +x and run it; it needs
glibc 2.35 or newer. If it will not start, the machine has no FUSE 2: install
libfuse2, or run the file with --appimage-extract-and-run, which unpacks it
instead of mounting it.

macOS - perfstudio-...-arm64.dmg, Apple silicon. Drag it to Applications,
then run this before opening it:

xattr -dr com.apple.quarantine /Applications/PerfStudio.app

It is signed ad-hoc but not notarized, so Gatekeeper blocks the first launch.
On macOS 15 and later the dialog it raises defaults to Move to Trash, which
deletes the install. Clearing the flag first avoids the dialog entirely. On an
Intel Mac, install from source.

Anywhere with Python 3.12+: pip install -U perfstudio, then perfstudio.
Same release, no installer and no SmartScreen warning; Qt and VTK are most of a
400 MB download the first time.

Checking your download - SHA256SUMS carries the SHA-256 of all three
files. sha256sum -c SHA256SUMS on Linux, shasum -a 256 -c SHA256SUMS on
macOS, Get-FileHash <file> in PowerShell. PerfStudio does this for you when it
fetches an update itself: Help - Check for Updates. It proves the file
arrived intact, not who built it - that second claim needs the code signing
this project has not bought yet.

This is pre-alpha software. Nobody has yet built a real board by following a
generated guide - check every measurement against the part in your hand.