fix(lex-knowledge): gap fixes — metadata→context, embed fallback, retire tag, error logging

[Esity]

$(cat <<'EOF'

Summary

Seven correctness fixes across runners/ingest.rb, runners/query.rb, runners/maintenance.rb, and actors/maintenance_runner.rb.

• Fix 1 — metadata: → context: in ingest_to_apollo (CRITICAL): chunk provenance (source_file, heading, section_path, chunk_index, token_count) was silently dropped by Ruby's double-splat because handle_ingest expects context:. Every lex-knowledge chunk stored '{}' in apollo_entries.source_context, breaking all maintenance queries that use source_context->>'source_file'.

• Fix 2 — llm_embed_available? now accepts single-embed providers: previously only checked for Legion::LLM.embed_batch; any LLM that only exposes Legion::LLM.embed caused all chunks to be stored without embeddings. build_embed_map now falls back to per-chunk embed calls when embed_batch is unavailable.

• Fix 3 — retire_file uses valid content type and context: key: changed content_type: 'document_retired' (unknown to lex-apollo, silently stored as :observation with no signal) to content_type: 'observation' with tags: [..., 'retired', ...] for an explicit retirement signal. Also switched metadata: → context: to match handle_ingest signature.

• Fix 4 — retrieve_chunks logs errors instead of swallowing them: the anonymous rescue => _e; [] pattern silently hid DB/connection errors, causing synthesize_answer to generate hallucinated answers with no context. Now logs at warn level.

• Fix 5 — synthesize_answer returns nil on LLM error: previously returned the error message string (e.g. "Error generating answer: connection refused") as the answer: field, which callers presented to users as a real answer. Now returns nil and logs at warn level.

• Fix 6 — MaintenanceRunner#enabled? works for monitors-only installs: enabled? previously returned false when corpus_path was nil, so the maintenance actor was never scheduled for installs that use monitors: config. Now returns true when Runners::Monitor.resolve_monitors.any?. args falls back to monitors.first[:path] in that case.

• Fix 7 — query_count uses correct action string: ApolloAccessLog records use action: 'query' (set by lex-apollo), not 'knowledge_query'. The old value always returned 0, making quality_report totals useless.

Test plan

• Ingest a corpus file and verify source_context in apollo_entries contains source_file, heading, section_path, chunk_index, token_count
• Configure an LLM provider that only exposes embed (not embed_batch) and confirm chunks are stored with embeddings
• Remove a file from a corpus and confirm the retirement entry has content_type: 'observation' and tags include 'retired'
• Simulate a retrieval failure (take Apollo offline) and confirm a warn-level log line is emitted, not a hallucinated answer
• Simulate an LLM failure during synthesis and confirm answer: is nil in the response, not an error string
• Configure a node with monitors: only (no corpus_path) and confirm MaintenanceRunner is enabled and scheduled
• Confirm quality_report total_queries returns a non-zero value after running queries
  EOF
  )