Parse through gcode-ast instead of the hand-rolled scanner

[sophiedeziel]

Swaps Parser.parseCommand to delegate to gcode-ast's parseLine + GCodeCommand compat view, then rewires the interpreter to dispatch on the typed AST nodes instead of lower-cased mnemonics.

Not for merge. Opened to make the tradeoff concrete.

Commit 1 — parse through gcode-ast

parseCommand delegates to parseLine; the local GCodeCommand class, GCodeParameters interface and isAlphaCode scanner are re-exported or deleted. Thumbnails, slicer detection and layer metadata are untouched.

Geometry is identical on all 7 demo fixtures (240 layers / 97,574 points on 3DBenchy, plus every JobStats field and the bounding box).

3DBenchy, parse → interpret:

	time	heap
current	69 ms	47 MB
this PR	182 ms	206 MB
if spans: false were reachable	153 ms	128 MB
if raw: false were too (loses src)	133 ms	104 MB

Two tokenizing regressions — both real gaps in gcode-ast, both with existing tests here:

• G 1 E 42 X 42 (spaces between letter and number) → gcode: 'g' instead of 'g1'
• G1 X1.2.3 Y5 → drops x entirely; we take the leading 1.2. Arguably better, still a behaviour change.

The perf knobs are unreachable. spans: false doesn't typecheck — toGCodeCommand requires a span. raw: false drops line.raw, and GCodeCommand.src is public API asserted by two tests.

GCodeCommand.node is most of the memory. The back-reference retains the whole typed node — words[], commandWord, spans — per line.

Commit 2 — typed dispatch

The registry is keyed on the AST command type, and each handler receives its node narrowed. Registering a handler for a type the parser cannot produce is now a compile error instead of an entry that never fires.

• probe re-derived the three G31 forms from a P word and an axis-word count — now reads command.form
• selectTool parsed the index out of the mnemonic — now reads command.index, so T8+ works instead of falling off a t0–t7 registry
• setInchUnits/setMillimeterUnits collapse into one setUnits; the node carries which unit it selects

Geometry is still identical on all 7 fixtures, and dispatch cost is neutral — the overhead is all in tokenizing. 873 / 875 tests pass (same 2 as commit 1). Coverage stays at 100%.

Two things worth review:

• G92 reads "bare" off words, not the typed fields. The node names only X/Y/Z/E, so G92 F3000 would look bare through them and reset a workspace it should leave alone — there is a test pinning exactly that.
• modal is deliberately unregistered. gcode-ast now emits a node for continuation lines like F140; registering a handler would make CNC files gain geometry they don't render today. Separate decision, separate PR.

Handler tests now build commands by parsing a real line instead of hand-pairing a mnemonic with a params bag, so they can no longer assert on a shape the parser would never produce.

Nothing here is public API — Interpreter, handlers and CommandHandler were never exported.

Still unsolved

Depends on github:xyz-tools/gcode-ast#add-prepare-script — gcode-ast#14, needed because a github: dep resolved to a package with no dist/. A real dependency means publishing to npm.

Verdict

The swap works, the typed dispatch is a clear win on its own, and the geometry proves both. Still not mergeable at ~2.6× time and ~4× memory, and the remaining fixes are all upstream: the two tokenizing gaps, an opt-out-friendly toGCodeCommand signature, a way to skip node, and npm.

Assisted by Claude Code - Opus 5