[psyinf]

Stacked on #19 (feature/to-ascii). Retarget to main once #17–#19 are merged.

Problem

A merge could not read a named graph input:

(plots: plots, ticks: in.ticks) -> GatedImmTracker -> tracks
2:25: expected an edge name in group but found '.'

The workaround was a pass-through stage to copy the input onto an ordinary edge. Plain in in a group was
parsed, but it was never registered as a graph input.

Change

• Parse: both parsers (lexy and the deprecated handwritten one) accept an optional .key after a group edge,
  including after name:. They record each edge's key and location.
• Build: buildStatement checks the group edges. in.<key> is spelled as its edge. x.key gets the usual
  ". only on in/out" diagnostic. out, out.<key> and end inside a group give "may only appear as the last
  term". Every in / in.<key> is registered through a new bindInput helper, the same one the single-source
  branch now uses, deduplicated by edge.
• Types: wireMergeSlots now infers the type of an undeclared input read only by a merge slot from that
  slot's type. Before, such an input stayed untyped. Declared inputs are checked as before
  (slot 2 of 'X' expects ... but the graph input 'in.quotes' carries ...).
• Docs and example: namedPipeline gets section 7, (greeting: in.greeting, name: proper) -> Greet, with its
  toAscii listing. README, EXAMPLE.md, CHANGELOG and the grammar comment in GraphLang.hpp describe the new form.

Tests

• GraphInputsTests:
  • positional, named and mixed groups over named inputs
  • plain in in a group of a single-input graph
  • declared-type checks and "never reads" with group-only readers
  • type inferred from the slot, with std::invalid_argument on a wrong type
  • must-fail cases for msg.x, out, out.x
  • GraphProgram::inputs and the input's location
• FilterGraphTests:
  • lexy and handwritten parsers agree on the new forms and diagnostics
  • toAscii roots a tree at a group-only input

All 124 tests pass locally (MSVC debug developer preset).