v3.2.0 — multi-profile reasoning mining; embeddings stop inheriting a stranger's endpoint

Added

• Reasoning mining can read more than one Claude-Code profile. New [reasoning_training] extra_transcript_dirs (also SURREAL_MEMORY_REASONING_EXTRA_DIRS, comma-separated) lists additional profile roots to scan alongside the implicit ~/.claude — each entry is the directory that holds projects/, e.g. ~/.claude-ZAI. Running a second profile for a different vendor's models is the normal way to do that on one machine, and until now every trace from that profile was invisible to mining no matter how mining_models was set. Roots are de-duplicated, missing ones are skipped rather than fatal, and each file's project attribution resolves against its own root. (#156)
• The injection map's target field is now a combobox. It offers every detected model while staying free-text, because targets are matched with fnmatch — a glob (glm-*) and a model that has not appeared in any transcript yet both have to remain typable. Targets are deliberately not filtered by has_thinking_text: that gate applies to sources, and a target is the recipient. (#156)

Fixed

• Embedding requests silently inherited an ambient OPENAI_BASE_URL. With no surreal-memory endpoint configured, the OpenAI SDK fell back to that environment variable — the ecosystem-standard knob for pointing some other tool at a proxy — so a brain's memory text was sent to a host the operator never configured for embeddings, failing later as a confusing error from a vendor that had never heard of the configured model. OpenAIEmbedding now always passes an explicit base_url, defaulting to https://api.openai.com/v1, and logs a warning when it ignores a set OPENAI_BASE_URL. (#156)
• The Stop hook's loopback privacy gate was defeated. It resolved the endpoint from config-or-env, checked it was loopback, and then built the provider without passing it — so the constructor re-resolved from the environment alone and an endpoint configured in config.toml cleared the gate while the client pointed elsewhere. The checked endpoint is now handed to the provider. The dashboard's "test embedding connection" endpoint had the same omission with no gate at all, and is fixed the same way. (#156)
• smem reindex hid the reason it failed and exited 0 anyway. It printed one identical batch N-M failed (skipped) line per batch, swallowing the actual exception, kept grinding through every remaining neuron repeating the same failure, and reported success to its caller. It now reports the real error once, aborts after a few consecutive failures when nothing has been embedded, and exits non-zero when a run embeds nothing. (#156)

Changed

• Breaking for one setup: if you relied on OPENAI_BASE_URL alone to point surreal-memory at a LiteLLM/vLLM/Azure/self-hosted OpenAI-compatible server, with provider = "openai" and no endpoint in either config.toml or SURREAL_MEMORY_EMBEDDING_ENDPOINT, embeddings will now go to api.openai.com instead. Set [embedding] endpoint (or SURREAL_MEMORY_EMBEDDING_ENDPOINT) to restore the old routing — the same one-key change the new warning names. This is why this is a MINOR release rather than a patch.

Full diff: v3.1.0...v3.2.0