[P1] Verifier: remove duplicate ev-node call and cache Celestia lookups per DA height #40

Priority: P1 (performance) · Effort: M

Problem

Per verification attempt (verifier.go:verifyAttempt):

• evnodeClient.GetBlock is called, then GetBlockWithBlobs, which calls GetBlock again (internal/clients/evnode/client.go:68). That's two identical ev-node RPCs.
• celestiaClient.GetBlockTimestamp(daHeight) runs once per block. ev-node batches many blocks into one DA height, so the same timestamp is fetched N times.
• Data path: GetBlobsAtHeight downloads all blobs at (daHeight, dataNS) once per block, so a 100-block batch downloads the same blobs 100 times. Header path: GetIDs once per block.

Proposed fix

1. Pass the already-fetched *GetBlockResponse into the blob-marshalling helper, or call only GetBlockWithBlobs. This saves one RPC per attempt.
2. Add a small memoization map in celestia.Client keyed by (daHeight, ns) for the timestamp, IDs and blobs, protected by a mutex and with a ~2 min TTL that's checked on insert. Use golang.org/x/sync/singleflight (already in go.mod) so concurrent workers share one in-flight fetch. Add a simplify: comment noting there's no LRU and the upgrade path.

Pros / Cons

• ✅ Cuts Celestia load and bandwidth by roughly the batch size (often 10-100×), lowers latency, and reduces timeouts on public nodes.
• ⚠️ Memory for cached blobs (up to ~2MB each), so keep the TTL short. Never cache negative or "not found" results; otherwise blocks that land later would fail.

Acceptance criteria

• Test: N blocks at the same DA height cause exactly one blob fetch.
• Test: a "not found" result is not cached.