epic(scrybe-mcp): reference resolution & grounding (resolve before the model)

[hartsock]

Part 2 of the Conversational Document Editing arc — see docs/design/vision-conversational-editing.md. Depends on the object-addressability epic (Part 1).

Goal

Make grounding Scrybe's job, not the model's. Resolve a reference to an object handle + revision + local context before the model is invoked — fast, deterministic, and cheap in tokens. The model never receives a whole document to search.

Surface

• document.resolve_reference("Figure 2") → {document_id, object_id, revision, geometry, contents, surrounding_context}.
• document.references() — enumerate the referenceable objects in a document.

Named vs. deictic (vision doc §4 — do not conflate)

• Named references ("Figure 2", "the installation section", "the performance table") — deterministic; this epic.
• Deixis ("this paragraph", "move this above", "that section") — resolved from the editor's current selection when app-driven (cursor = "this"): a real near-term partial win. Full multimodal deixis (pointer, eye-tracking, spatial context) is deferred to a later epic gated on a multimodal input layer. In headless/no-selection mode, "this" returns an ask, not a guess.

Depends on

• Part 1 (object addressability), epic: MCP rebuild — shared scrybe-tools ToolSpec registry #122 (MCP rebuild).

Acceptance

• Named references resolve deterministically to an object handle with no full-document search.
• App-driven deixis resolves "this" from the live selection.
• Headless / no-selection deixis returns a disambiguation prompt rather than guessing.

Milestone: ~v0.8–v0.9.