perf: cache journal entries, parallel loading, boot warmup for search

[Ghilteras]

Problem

Every search re-reads and re-parses every entry and loads embeddings serially. With a growing journal, each search pays full O(N) disk I/O plus cold model load after each process start.

Solution

• Add a per-store fingerprinted cache for parsed journal entries and plain-array sidecars.
• Stat entry and sidecar files in parallel, load cache misses in parallel, and gracefully skip malformed entries.
• Warm the embedding model at boot after configuration is read and only inside the journal-enabled feature branch.

Not included

• Title floor (PR feat: title-anchor floor for journal semantic search #31).
• Versioned embedding format or multilingual model; these are out of scope for upstream.

Tested

• bun test (29 pass, 0 fail)
• bun run typecheck (exit 0)

This ships in our fork since v0.4.1, with one hardening fix from review: warmup must run after config and inside the feature branch. The transformers.js-style singleton is first-init-wins.