Revisit the anonymous-snippet diagnostic label in Runner.render(source:)

[leogdion]

Background

Following #163, Runner.render(source:originalPath:) made originalPath optional
(String? = nil) so a developer can render a bare string of SyntaxKit DSL without inventing
a path. originalPath is purely a diagnostic label — it feeds the #sourceLocation(file:)
directive in WrappedSource and the stderr path-rewrite in processFile; nothing is opened,
and the output cache keys off the source bytes (not the path), so the label never affects
caching or the rendered output.

When originalPath is nil, processFile currently falls back to a hardcoded literal:

// Runner.processFile
let absoluteInputPath =
  originalPath.map { URL(fileURLWithPath: $0).standardizedFileURL.path }
  ?? "source.swift"

This works and is covered by tests, but the fallback was chosen quickly and deserves a
second look.

Questions to revisit

• Magic string. "source.swift" is an inline literal. Should it be a named constant
  (e.g. WrappedSource.defaultOriginalPath, its semantic owner) so the label has one home
  and is discoverable? (An earlier attempt to add this constant was deferred in favor of just
  making the parameter optional.)
• Best fallback behavior. Options when no path is given:
  1. Synthetic filename label — current ("source.swift"), so diagnostics read
     source.swift:3:5: error: ….
  2. A clearly-synthetic marker like "<source>" / "<in-memory>" (signals "not a real
     file", but verify #sourceLocation + any downstream tooling tolerate angle brackets).
  3. Omit #sourceLocation entirely for anonymous snippets — most "honest", but then
     diagnostics reference the internal temp wrapper (Input.wrapped.swift) with wrapper-
     relative line numbers, which is likely worse for users.
• Label collisions. Two distinct anonymous snippets both report source.swift in their
  diagnostics. Harmless today (no cache collision — keyed on source bytes), but worth a note
  in the docs, or an optional caller-supplied label could disambiguate.
• Batch parity. The batch entry point render(sources: [RenderInput]) requires a
  RenderInput.url. Should there be an ergonomic path for rendering a batch of anonymous
  in-memory snippets too (optional/derived identifiers), or is that out of scope?

Acceptance

Decide and document the intended behavior for the no-path case; if a named constant or a
different fallback is chosen, update Runner.processFile / WrappedSource and the tests
(RunnerRenderFailureTests, RunnerBatchRenderTests) accordingly.

Context: implemented on branch feature/163-filesystem-free-render.