v0.10.5 — Load-Time Template Validation & Friendly Local-Model Errors

Direct response to production feedback (华能电厂 hypergraph KA, 840 docs): a template-config error that used to surface as an unrelated crash now fails at load time with the exact cause named.

🛡️ Load-Time Identifiers Validation

A template whose identifiers reference fields not declared in output (e.g. relation_id: '{name}|{type}' while relations.fields omits type) previously loaded fine and crashed at first merge with ValueError: max() arg is an empty sequence — zero connection to the root cause.

Now Template.create fails immediately:

ValueError: relation_id references '{type}' but relations.fields does not declare 'type'

Covers node_id, relation_id placeholders, relation_members, and time_field/location_field, across all five graph-family types.

🧠 Friendly Local-Model Errors

Without a config file, Template.create silently fell back to OpenAI defaults (gpt-4o-mini) — vllm/ollama users hit a cryptic auth error far from the cause. The error now lists all three ways forward: he config init, -p vllm local setup, or passing clients explicitly.

🔒 ontomem 0.6.1 (companion release)

• KeyExtractionError: 100% key-extraction failure now raises an aggregate error (count + distinct causes) instead of silently returning empty
• key_extraction_failed warnings aggregate per distinct error (x118), no more alarm floods
• remerge="mechanical": exact rollbacks on large shared keys use the deterministic field merge (seconds, no LLM) — fixes the 35+ min hang reported with runaway LLM generation on batch-built KAs; auto mode falls back automatically above 12 survivors per key

Full changelog: v0.10.4...v0.10.5