Mesh and NRP

Picks one claimant for a namespace: skips itself, stale and pending entries, prefers the best score, honors name and tag selectors, can try a runner-up, and falls back to the root space.

The same selection returning a ranked, capped list of candidates.

/.mesh/resolve/multi returns scored candidates, sorted, with max, min_score and staleness limits, and leaves out itself.

POST /.mesh/announce: validation, throttling, signed versus pending entries, namespace squatting, first-key-wins, and the trusted-key allowlist.

The index of known monads: read and write, order by last seen, lookups by namespace or name, the root identity hash.

A claimant's score: fresher, more resonant and faster scores higher; always finite and in [0, 1]; deterministic; shifts with forward results.

Custom scorers registered at runtime (patches): storage, operations, and their effect on scores.

Scorer weights start at defaults, move with rewards, and never fall below a floor.

The learning step end to end: exact weight update, weight floor, reward formula, convergence over many requests.

Health signals on learned weights: stable, dominant scorer, dead scorer, oscillation, no learning.

Each routing decision is recorded and its outcome becomes a reward (a fast success scores near 1), written as JSONL.

The same decision log and reward for synthesis decisions, and how they feed weight updates.

Combining answers from several sources: all agree, a minority winner, or contested; failures, divergence and disclosure.

Asks several monads and returns one public value on quorum, or a contested null; never forwards to pending entries or the cloud metadata address.

Parses a bridge target (namespace:op/path) and the monad[frank] path syntax.

/nrp never sends 'stealth' on the wire: a namespace the kernel does not confirm resolves to 'closed'.

/nrp serves reads under a declared root and refuses a single label nobody declared.