[psyinf]

Implements TODO item "Several named graph inputs". Stacked on #15 — merge that first; this PR then retargets to main.

What

A graph can read named inputs, mirroring GraphOutputs:

DslFilterGraph<GraphInputs, Match> matcher(R"dsl(
    in.orders -> ParseOrder -> order
    in.quotes -> ParseQuote -> quote
    (order, quote) -> Match -> out
)dsl",
    GraphInputs::of<Order, Quote>("orders", "quotes")); // optional

matcher.push("quotes", Quote{...});                                       // one run, one input
matcher.filter(GraphInputs{}.set("orders", order).set("quotes", quote));  // one run, several inputs

Semantics

• One push / filter is one run. Inputs not fed are empty, exactly like dropped paths: readers are skipped, merges see holes. No new rules for merges or outputs.
• GraphInputs::of<Types...>(keys...) declares input types, checked at construction like any edge. Declared-but-unread and read-but-undeclared inputs are diagnostics.
• Without a declaration, an input is typed by its first reader (diagnostics say so: the graph input 'in.orders' (typed by its reader 'ParseOrder')).
• Feeding an unknown key throws std::out_of_range; a mistyped value throws std::invalid_argument.
• in with GraphInputs, or in.<key> with a plain input type, is a diagnostic. DslFilterGraph<In, Out> with a plain type is unchanged.

Also

• .<key> on anything other than in/out (e.g. msg.x) is now a diagnostic; it used to be silently ignored.
• dsl::GraphProgram::inputs lists the inputs a program reads; validateDslGraph(text, inputTypes) overload.
• Internally, edge types live in a small EdgeTypes helper, and GraphPlan::run takes one value per input.

Docs / tests

README ("Several named inputs"), EXAMPLE.md section 5 + textPipeline demo 4b, CHANGELOG, TODO (item removed, rest renumbered). New GraphInputsTests.cpp (11 cases, incl. nesting a named-input graph as a stage) plus parser-agreement cases. All 107 tests pass (MSVC debug, ASan).