chore: Implement openranking

[dcadenas]

Summary

nos-followers already computes PageRank, follow-graph counts, a trust verdict, and similarity recommendations, but only exposed them through its own /api/v1.
This change adds the part of the Open Ranking protocol that fits our graph, so any Open Ranking client can read that data without a custom integration.
The new endpoints are additive, so /api/v1 is unchanged.

What Changed

New top-level JSON-over-HTTP endpoints, discoverable from a capability document at GET /.well-known/open-ranking:

• POST /rank ranks pubkeys by global PageRank, returning a rank for every requested pubkey (unknown pubkeys rank 0).
• POST /reputation returns PageRank as the rank, follower and following counts, and our trust verdict in the provider x extension.
• POST /followers returns the top followers of a pubkey by PageRank, plus the total follower count from the follow edges.
• POST /recommend returns personalised for-you recommendations (a pov is required), up to 10.

Input is validated per the spec (400 malformed, 413 too large, 422 semantic) with an X-Reason header, and the routes share the existing CORS, timeout, and rate-limit layers.
Two batched Neo4j reads were added (get_pageranks and get_top_followers); the rest reuses existing repo methods.

/recommend never blocks on the heavy similarity query.
A cache miss starts a background compute through the shared recommendation queue and returns an empty list with a short ttl, and the client retries once it is warm.
A transient compute failure now self-heals: the queue treats a cached failure as a miss and re-enqueues on the next request, instead of wedging the endpoint on a 500 until the status cache expired.
New Prometheus metrics track the cold-miss rate and the warm-up compute time, and CI now gates rustfmt and clippy.

Testing

I rebased onto main and ran the suite on the pinned 1.84.0 toolchain.

• cargo test --all-features --features ci: 86 passed, 0 failed
• cargo clippy --all-targets --all-features -- -D warnings: clean
• cargo fmt --all -- --check: clean

The two new repo queries also have Neo4j integration tests, gated out of CI and run locally against the compose db service.

Risks

The change is additive, so existing /api/v1 consumers are unaffected.
For a cold account, /recommend returns an empty list first and the real list once the background compute finishes, so clients retry after the ttl.
The warm-up histogram measures the compute time, which is the dominant cost, not the time a request waits in the queue.