Known rough edges

Small defects and unfinished corners that are worth knowing about before working near them, and that are too minor to have their own page. Each was checked against the code, not inherited from an older list. Design-level limitations of the parameter system are in parameterSystem.md, under Known Limitations, instead.

Delete an entry when it is fixed. A list that keeps its history stops being read.

Parameters

• *ParameterAnalysis.chainActionsOnSheet picks the wrong declarer.* Set.minElement declarers is the alphabetically first sheet that declares the parameter, not the outermost one. The default value and the description copied onto the intermediate sheets of a materialised bind-to-top chain therefore come from an arbitrary sheet. Worse, BindOffer.BindsTo prefers the top sheet where it declares the name, so the button can say "connect this to TOP.width" while copying the default and description from a different sheet entirely.
• Deleting a parameter checks slots but not other parameters. ParameterView.deleteParameterBox lists the ParamSlots that use the parameter and refuses while any remain, but does not look in DefaultBindings, so a parameter defined in terms of another can be orphaned. Only reachable through a hand-edited file today, since the UI writes only PInt defaults — but editParameterBox handles expression-valued defaults, so the system claims to support them.
• *ReloadSelectedComponent is fed the wrong number.* For a BusSelection LSB or a BusCompare value it receives that number as "most recent bus width", which is what the properties pane then offers as the default width for the next component.

Wire routing and separation

How these two passes are meant to work is in wireRouting.md.

• *string on an [<Erase>] id means two different things.* InputPortId, OutputPortId and friends are erased by Fable, so string portId is the bare id in the app and InputPortId "…" under .NET. BusWireRoute used it for five model.Symbol.Ports lookups, which therefore threw in any .NET test that routed a wire; those are now inputPortStr/outputPortStr. Nothing checks for the pattern, and the same trap is open wherever an erased id meets string, sprintf or an interpolation.
• *removeWireSpikes and removeModelSpikes have no callers.* Spike removal is written, exported and never run: separateAndOrderModelSegments ends with removeModelCorners and nothing calls the spike pass. Either the artifact it removes no longer occurs, in which case delete it, or it does and the pass was dropped by accident.
• *hasOverlap and hasNearOverlap each have a clause that can only fire on exact equality.* Both write b1.MinB where the third argument should be b1.MaxB. Harmless — the other two clauses already decide overlap in every case — but it reads as if it were load-bearing.
• *makeClusters calls the head of a descending-sorted list lowestLoc2Index.* It is the highest index, so the test that follows it is not the "did the downward search fail to reach the starting segment" check it appears to be. It errs towards the branch that splits off a second cluster, which is the safe one.
• Zero-length segments turn up in the middle of wires. They belong beside a nub, where they make the first visible segment draggable, and nowhere else: a zero segment mid-wire is two coincident vertices - a vertex that is not a vertex - and a separation move which crosses one draws the wire back over itself. Redrawing reg16x8 from the 3cpu fixture leaves 3 of them and one such spike, with or without same-net branching, so ordinary routing and separation produce them too. removeWireSpikes was written to clean up after this and is called from nowhere; removing the cause is better than calling it.
• Dead code kept alive. snapToNet (and copySegments, generateEndSegments, which serve only it) was the first attempt at what sameNetRoutes now does, and is still there and still unreachable — it only ever handled 5 or 7 segment unrotated wires and copied from whichever wire of the net came first out of a Map. It should go. expandCluster computes lowestDownwardsIndex for a guard that is commented out. adjustSegmentsInModel binds Option.get line.Seg1 and never uses it. The doc comment on Constants.separateCaptureOverlap describes maxCornerSize.

Component libraries

• A typed library name becomes a directory name unchecked. MiscMenuView.saveAsLibraryComponent joins the user's text onto the libraries directory with no validation beyond non-empty, so a name containing a path separator or .. writes outside the intended directory. Sheet names are validated by maybeWarning; library names are not.
• Name collisions have no rule. CustomComponentType.Name refers to sheets by name, so a library Adder and a user Adder cannot coexist in one project. Namespacing or rename-on-import is the missing piece.
• Sweeping a library sheet leaves its backups. MenuHelpers.sweepUnusedLibrarySheets removes the .dgm when the last instance goes, but the sheet's files under backup/ stay.
• A copied-in library sheet does not track the library. Once materialised it is an ordinary project sheet and never sees a later version of the library component. This is probably what is wanted, but it is nowhere stated to the user.

Simulation

• The memory guard that bounds how many cycles a simulation may run is now too conservative. SimTypes.Constants.heapBytesPerComponent is 1000 + 350 * ports, measured when the waveform simulator described every wave the simulation offered and kept the lot in its model. It no longer does - it keeps records only for the waves selected - so roughly 260 of those 350 bytes per port is a cost that has gone, and designs are being refused that there is memory for. The number was measured and its replacement has not been; the comment beside it says how to remeasure (build phase deltas under --log=perf, and window.issieDev.simStats() for the counts to divide by).

Refactoring worth doing

• SheetLayout.saveSheet and sheetBody are the same function twice, differing only in saveStateToFile versus stateToJsonString; both also evaluate paramDefsOf a second time after toCanvasState has already done it.
• "The custom components named in a canvas" is written three times: ComponentLibraries.customSheetsUsedBy, the same List.choose inline in ComponentLibraries.unusedLibrarySheets, and SheetLayout.saveLibraryComponent's requiredBy.
• LibraryHeader is built field by field in both MiscMenuView.saveAsLibraryComponent and SheetLayout.saveLibraryComponent. The format is versioned, so the two writers drifting matters.
• MiscMenuView.maybeWarning calls ComponentLibraries.reservedPrefixOf twice and then uses .IsSome/.Value, against the Option-throughout convention.
• FilesIO.modifiedTimeMs has no callers. It was written for a library-index scheme that was dropped in favour of lazy reading.
• VerilogComponent/TestParser.fs (937 lines) is a hand-driven test runner that compiles into the shipped renderer, reads paths that exist only in a dev checkout, and keeps printf on the SourceHygiene allowlist. It should go once its corpus runner is in Expecto — see verilogTesting.md, under Route A.

Documentation

• The screenshots on the documentation site are a version behind. docs/img/homePage/* and docs/img/userGuide/* show a Sheets menu and an Electron application menu bar that no longer exist — the bar is now Project | Sheet | Edit | View, drawn inside the app. The workflows are unchanged, so they are still usable. scripts/inspect-canvas.js can screenshot the running app, so a refresh can be scripted rather than done by hand.