🦌 Add speculative PR metadata generation to hide Claude latency

[zdavison]

Summary

Adds an experimental speculative_close feature that starts generating PR metadata (branch name, title, body) in the background while the user is reading the post-session menu. On the happy path where the user picks "p" (create PR), the metadata is already ready or nearly ready, hiding most of the Claude inference latency.

If the user picks any other option, the speculative work is discarded at the cost of one extra Claude API call.

Changes

• packages/deerbox/src/config.ts — adds experimental.speculativeClose config option with TOML and repo-local parsing support
• packages/deerbox/src/cli.ts — passes speculativeClose from config into the post-session context
• packages/deerbox/src/post-session.ts — kicks off speculative fetch and metadata generation before prompting the user; passes result to createPR; also fixes destroy() to use .catch(() => {}) instead of awaiting
• packages/shared/src/git/finalize.ts — exports generatePRMetadata and PRMetadata; adds optional fetchDone param to reuse a prefetch; parallelises the three git reads after fetch using Promise.all; createPullRequest accepts and uses speculativeMetadata
• packages/deerbox/src/git/finalize.ts — re-exports generatePRMetadata and PRMetadata from shared
• packages/shared/src/index.ts — exports generatePRMetadata and PRMetadata from the shared package

---

Created by deer — review carefully.

[zdavison]

why does this incur an extra claude api call for other options?
list each option and what happens

also, speculativeClose should also work for m because commit message is generated with claude, and its used for both m and p

[zdavison]

Is this going to cause problems? the flow before was stageAndCommit -> generate right? now it'S generating before stageAndCommit?