Cache worktree reconciliation fetches and PR lookups across pages

[chubes4]

Problem

Paged metadata reconciliation in DMC v0.33.0 prevents one giant timeout, but each page can still redo expensive repo fetches, git probes, and GitHub PR lookups.

On a large workspace (~580 worktrees), page-by-page reconciliation is still slower than it should be. The first page found useful cleanup eligibility promotions, but draining the inventory remained operationally slow.

Desired outcome

Persist/cache reconciliation probe results across pages/runs where safe.

Acceptance criteria

• Repo fetch/prune happens at most once per repo per reconciliation run, not once per page whenever avoidable.
• GitHub PR lookup results are cached by repo/PR number with a sensible TTL.
• Cached evidence is included in output so operators can tell whether a decision came from a fresh probe or cache.
• Cache invalidation is conservative: dirty/unpushed/current worktree safety checks still run fresh at apply time.
• Works for direct apply, job-backed drain, and time-budgeted drain modes.

Related

• Scale worktree cleanup and metadata reconciliation for large inventories #333
• Repair and prevent missing lifecycle metadata on worktrees #334
• Automatically mark worktrees cleanup-eligible from PR and branch lifecycle signals #336

AI assistance

• AI assistance: Yes
• Tool(s): OpenCode (GPT-5.5)
• Used for: Drafting issue from post-release deployment and cleanup attempt evidence reviewed by Chris.