feat(Syntax/Anaphora): add Anaphor + Diagnostic substrate; JSON-backed Landau2026

[hawkrobe]

Extracts Hankamer & Sag's deep/surface anaphora theory ([hankamer-sag-1976]) as a reusable, framework-neutral substrate, and reworks Landau 2026 (the EIR ellipsis test) onto it.

• Syntax/Anaphora/Basic.lean — root Anaphor namespace: Depth (deep/surface) + Control (pragmatic/syntactic, H&S's primary diagnostic) + Depth.HasInternalStructure/control with the allowsPragmaticControl ↔ ¬hasInternalStructure law + the HasDepth capability (depth-axis analogue of Bound). Orthogonal sibling to the binding-theoretic Bound/Features.BindingClass axis.
• Syntax/Anaphora/Diagnostic.lean — a general, polymorphic Diagnostic Ω C: abductive test where each outcome is consistent with a Set of latent values; Conclusive/SoundFor/Decides/Refines + the ofCauses builder, with Depth-specialised DepthCause/testOutcome.
• Studies/Landau2026.lean regrounded — the EIR prediction is the substrate structural property (was three renamed identity functions); EIR and extraction are Diagnostic values, and the headline eir_decides : eir.Decides Depth.testOutcome says EIR decides surface-vs-deep where extraction can't. Data moved to Data/Examples/Landau2026.json (the paper's (18a)–(47b) stimuli, PDF-verified); the study projects classifications off paperFeatures/judgment. 434 → 223 lines.
• Studies/Steedman2000.lean — second HasDepth consumer; surface_not_syntacticallyMediated surfaces a CCG-vs-H&S non-alignment (VP-ellipsis is surface yet CCG-non-mediated).