11.1.1 — AICF dispatch mode, swarm-served embeddings

Node: AICF dispatch mode (ANIMICA_AICF_DISPATCH, on by default in the docker entrypoint)

• Jobs carry a spec.kind (chat default, embed, classify, batch). Deterministic kinds run K=1 and are only offered to workers that advertise them in workerRegister hardware.kinds; chat keeps racing K workers for quality. 1-token probes are K=1 — they were 90% of job volume, each raced 3-way.
• Idle/fast-worker routing with a fairness share cap, short K=1 leases, live in-flight accounting, quality-floor widening to K=2.
• New RPC aicf.workerReleaseClaim {address, job_id} — a worker that can't answer hands its claim back instead of pinning it for the lease.
• Additive schema; byte-identical behaviour with the flag off.

Embeddings served by the worker swarm

• POST https://animica.dev/v1/embeddings (OpenAI shape; model animica-embed → bge-small-en-v1.5, 384-d) plus /v1/embed and /v1/healthz. Results travel as one EMB1 line (base64 float16 + sha256) the bridge verifies; falls back to local MiniLM and the response model names the real producer. /x402/embed now rides the swarm.
• Browser worker (pool.animica.org/serve) and the Animica Serve Android wallet (0.8.1) serve embed jobs.

Fixes

• Setup advisor no longer keeps the network "warm" with a 1-token job every 70 s (reads aicf.workerCount instead; ANIMICA_ADVISOR_WARM_PING=1 restores it).
• Restored code that shipped in 11.1.0 but lived only in the working tree (MCP paid tools/credits, pool inference-wallet stat, pool block-share metric, release-gate artifact checks) and committed it.

pip install -U animica