feat: mine every profile's reasoning, pick injection targets, stop leaking embeddings

[acidkill]

Summary

Three related fixes, all found while investigating a report that reasoning mining "only offers Anthropic models".

1. Mining now sees every Claude Code profile

reasoning_miner resolved exactly one root — ~/.claude — so a machine running a second profile (the normal way to use a different vendor's models on one box) had every trace from that profile invisible, and no mining_models setting could reveal it.

New [reasoning_training] extra_transcript_dirs (also SURREAL_MEMORY_REASONING_EXTRA_DIRS) lists additional profile roots — each entry is the directory that holds projects/, e.g. ~/.claude-ZAI.

Two traps handled deliberately:

• Project attribution. _project_from_path resolves against a projects_dir. Flattening per-root discovery without carrying the owning root would silently mis-label one profile's traces as the other's. The project name is now resolved at discovery time, while the root is still known — with a regression test that puts a same-named project dir in both roots.
• Double counting. Roots are de-duplicated after resolution, so listing ~/.claude again as an extra can't scan it twice. Scan-state keys are absolute paths, so two roots can't collide there.

Verified against a real second profile: 150 glm-5.2 traces became visible, alongside the existing Anthropic ones.

2. Injection target is a combobox, not a text box

The source was a dropdown; the target was a bare <input type="text">.

It's now a combobox (<datalist>) offering every detected_models entry while remaining free-text, because:

• targets are matched with fnmatch, so a glob (glm-*) is a legitimate value a plain <select> would have taken away;
• a model that has never appeared in a transcript still has to be selectable.

Targets are deliberately not filtered by has_thinking_text — that gate applies to sources; a target is the recipient and may emit no thinking of its own.

3. Embeddings stopped inheriting a stranger's endpoint

With no surreal-memory endpoint configured, OpenAIEmbedding omitted base_url entirely, so the OpenAI SDK fell back to the ambient OPENAI_BASE_URL — the ecosystem-standard knob for pointing some other tool at a proxy. Result: a brain's memory text sent to a host the operator never configured for embeddings, surfacing only as a confusing error from a vendor that had never heard of the configured model.

• OpenAIEmbedding now always passes an explicit base_url (default https://api.openai.com/v1) and warns when it ignores a set OPENAI_BASE_URL.
• The Stop hook's loopback privacy gate was defeated by the same omission: it resolved the endpoint from config-or-env, checked it was loopback, then built the provider without passing it, so the constructor re-resolved from the env var alone. An endpoint configured in config.toml therefore cleared the gate while the client pointed elsewhere. It now hands over the checked endpoint.
• The dashboard's "test embedding connection" endpoint had the same omission with no gate at all — fixed the same way.

4. smem reindex stopped hiding why it failed

It printed one identical batch N-M failed (skipped) line per batch with the actual exception swallowed, ground through every remaining neuron repeating the same failure, and exited 0 regardless. This cost two full diagnostic rounds in practice. It now reports the real error once, aborts after a few consecutive failures when nothing has embedded, and exits non-zero on a total wash-out.

Deliberately rejected

A reviewer pass suggested also changing _create_provider to pass base_url=endpoint instead of endpoint or None. Verified as a no-op ("" or env ≡ None or env) and dropped rather than adding churn.

Breaking change (one setup)

If you relied on OPENAI_BASE_URL alone to point surreal-memory at a LiteLLM/vLLM/Azure/self-hosted OpenAI-compatible server — provider = "openai" with no endpoint in either config.toml or SURREAL_MEMORY_EMBEDDING_ENDPOINT — embeddings now go to api.openai.com instead. Set [embedding] endpoint (or the env var) to restore the old routing; that's the same one-key change the new warning names. This is why the bump is MINOR, not PATCH.

Test plan

• Lint, format, mypy (351 files), check_dead_modules.py, security scan — all clean
• Full non-stress suite: 6817 passed (only the known transient live-SurrealDB/xdist errors, confirmed passing in isolation)
• Coverage 70.64% (gate 67%)
• mkdocs build --strict clean; all four doc generators re-run
• Dashboard: tsc -b and eslint clean; bundle rebuilt (it is committed)
• Regression tests proven to fail before the fix: reverted _ensure_client to the old shape and confirmed 2 of the 5 new isolation tests fail, then pass again after restoring — a test that cannot fail proves nothing
• New multi-root tests cover: second root discovered, per-root project attribution with a colliding project name, claude_dir override still suppressing extras, missing root skipped, duplicate root not double-counted, async ingest path
• New reindex tests cover: real error reported exactly once, early abort, non-zero exit on total failure, partial success still exits 0
• Backward-compat asserted explicitly: plain-OpenAI user (key only), env-var-only user, and the OpenRouter subclass all keep their existing base URLs