Cache graph and worktree identity within large proxy plans #212

The Microfreak B2 proxy refresh in Gitpa PR #24 plans 741 source pages and spends several minutes planning before atomic apply. A read-only inspection found repeated graph scans and Git identity subprocesses inside a single shared plan.

In mise-tasks/logseq/entity/proxy/page/lib/core.py, resolve_page (lines 160–171) scans every source page on each lookup, even when a raw or encoded candidate exists. The current graph has about 6,648 source pages. add_page, embedded-page traversal, and task_imports.py repeatedly resolve pages. graph_name, infer_codeforge, and validate_codeforge also repeat Git repository/remote/branch identity queries per page.

Gitpa's scripts/sync_namespace.py already creates one shared plan; splitting or rebuilding plans is not the cause.

Implement plan-scoped page indexes and repository identity snapshots:

• Preserve logical-name ambiguity detection, symlink validation, explicit source/destination worktree identity, and local-edit checks.
• Reuse the source/destination name indexes and Git identity within a plan; consider caching entity-owned properties in the same scope.
• Preserve atomic apply and source-expected hash verification. Do not cache across separate runs.
• Add meaningful counter tests proving that graph scans and Git identity calls stay bounded per graph, alongside existing proxy/worktree tests.
• Confirm planned bytes and source/destination identity behavior are unchanged, and record before/after timing on a namespace-sized fixture or real dry run.

Related to guide-image migration #204 and Gitpa PR #24. This is an independent follow-up, not a blocker for the B2 migration. Priority: normal. No dedicated implementation agent is running; no branch or worktree has been created for this issue. Discovery session: 01a110f1-45c5-7633-ad92-c6f9f7f131d8, read-only reviewer episode_sources now idle.

--AI(gpt-6-unspecified proxy planning performance)