docs: manual row identities are not discoverable by project-scoped pull #409

Problem

At ce30de7, manual document inserts use random row IDs, while docs pull selects deterministic project/scope/document prefixes. A successful manual set can therefore create a document that never materializes on pull. The missing-row set path also fails to forward the selected branch scope.

Reproduction

Use the actual setDoc and pullDocs functions with a SQL-aware in-memory query adapter and an isolated output directory. Set a document for a project and branch, then pull that view. Unchanged source stores a UUID-backed/main-routed row and the prefix-scoped pull misses it. Seed an older UUID row and repeat set/edit/archive to exercise migration and existing-row behavior.

Focused remediation

Use the deterministic composite row identity for manual inserts, select the exact project/scope throughout writes, forward the branch scope and reconcile UUID-era duplicates without resetting the selected version counter or creation timestamp. Foreign-project and sibling-scope rows must remain untouched.

I have prepared the implementation with original negative controls, 98 focused lifecycle tests, 480 neighboring docs/CLI tests, adversarial SQL review and an exact-commit Linux full-suite gate. Reconciliation uses bounded DELETE then UPDATE, not a distributed transaction. I will link the PR here.