BrainPalace 26.6.17

Fixed

• An interrupted index no longer leaves orphan chunks with a 0-document
  status. The folder record + manifest were written only at the very end of the
  pipeline, after the slow, rebuildable BM25 and graph steps. If the job was
  killed during that tail (e.g. the timeout-wedge above, or a crash), the chunks
  were already in the store but no manifest mapped them — so total_documents
  read 0, the file watcher had no folder to watch, and the graph was lost. The
  folder record + manifest are now persisted immediately after the chunk
  upsert, before BM25/graph, so an interrupt leaves the store and manifest
  consistent (document count + watcher intact, next run diffs incrementally) and
  only the derived graph needs a rebuild. (brainpalace-server)
• sentence_transformers (and its heavy ML stack) is no longer imported at
  module load. The CrossEncoder reranker did a module-top
  from sentence_transformers import CrossEncoder, so importing the reranker
  package — and therefore the whole services package and test collection —
  hard-failed in any environment without the dependency installed. The import is
  now deferred into the model-load path (it was already lazily loaded), so the
  module imports cleanly and the dependency is only required when a reranker is
  actually used. (brainpalace-server)
• OpenAI-SDK provider clients now use a bounded request timeout, so a dropped
  connection can no longer wedge an index job. Every AsyncOpenAI(...) client
  (OpenAI + Ollama embedding, OpenAI + Grok + Ollama summarization) was built
  without an explicit timeout, inheriting the SDK's 600s read timeout × 2
  retries. On a flaky link a half-dead connection left the indexing worker
  blocked inside await ...create(...) for up to ~30 min — and because job
  cancellation is cooperative (checked between chunks), the job became
  uninterruptible: --cancel was ignored and graceful shutdown timed out,
  forcing a SIGKILL. Clients now default to a 60s request timeout for cloud
  providers (OpenAI, Grok) and 120s for local Ollama, both overridable via
  config.params (timeout, max_retries). (brainpalace-server)