docs: COLLAB_DESIGN.md — loro CRDT integration design

[juanma-dev]

Summary

Design-before-code for the biggest remaining roadmap item, in the same spirit as DESIGN_NOTES.md. No implementation in this PR — the document exists to de-risk a ~4–5 week / ~4k LOC build before the first line is written, whether it's built by the maintainer or picked up by a contributor.

Key decisions drafted for review:

• §3 Architecture: dual-model op bridge. EditorState remains the local source of truth; a shadow LoroDoc owned by a CollabSession is the convergence authority. Local Steps translate to loro ops; remote loro diffs come back as history-exempt Steps through the normal transaction pipeline — adapters, history, and plugins untouched.
• §4 Mapping. Block structure → LoroTree (movable-tree convergence for splits/joins/wraps); textblock content → LoroText with Peritext mark ranges; a per-Step translation table is the implementation contract.
• §5 History. Remote edits never enter local undo; map_against turns out unnecessary for convergence under this architecture (loro merges; we never rebase) — the Phase-2 trait bet held, just not the way we expected.
• §6 Surface. Amends the v0.1 plan: a new taino-edit-collab crate rather than a collab feature on core (dependency discipline, public-API dogfooding, versioning freedom); the umbrella's reserved collab feature still delivers the original UX.
• §7 Testing. Deterministic host-side convergence simulation (scripted scenarios + seeded property tests) — no browser in the loop, which our DOM-free core makes possible.
• §8 Open questions gate the first implementation PR (tree granularity for atoms, initial-doc seeding, undo route, loro MSRV vs our 1.80 pin, umbrella feature shape).

🤖 Generated with Claude Code