v0.32.0

The first public release of the agents library: clients for model providers, and the KB
read tools a model calls, outside the engine.

com.vaelii/vaelii-agents is an Apache-2.0 library on the SSPL engine. It depends on
core; core never depends on it, and the engine holds no model- or provider-specific code.

• A Provider protocol and three backends: the Messages API, Ollama, and a scripted
  offline stub. VAELII_LLM_PROVIDER selects one, the stub is the default, and a backend
  that cannot be reached falls back to the stub instead of throwing. A caller's loop is
  written once against request and response maps that name no vendor field.
• KB read tools for a model, generated from the engine daemon's op table. Every tool is
  a read the daemon serves, and none writes. A tool call is held to the daemon's ceilings:
  a depth or a clock over one is refused :over-ceiling, and a count the engine builds
  whole is held to 100.
• A majority yes/no vote across Ollama hosts (VAELII_OLLAMA_HOSTS). A refusal, a
  truncation or an answer that opens with neither yes nor no is not counted as a vote.
• Each HTTP turn has one deadline, which bounds the whole turn, the response body
  included.

A limit under core 0.32.0. kb_escalate, kb_lookup, kb_query, kb_query_p,
kb_query_status and kb_argue run under no clock when a call names no :max-ms, and
kb_why_not, kb_search_tree, kb_compare_tacticians, kb_ask_within and
kb_prove_within do when a call sends no opts. Core's develop branch holds all eleven
to the daemon's clock ceiling, and this library's CI runs against it.

The number. This is a first release numbered 0.32.0: the engine, the plugin, the
adapters and this library ship one version string, so it requires core 0.32.0.

Full detail:
CHANGELOG.md and
docs/llm.md.