BrainPalace 26.6.32

Added

• feat(config): layered config resolution — code < global < project. The
  server now resolves each config key by merging the global XDG config.yaml
  under the project config.yaml (load_merged_config_dict / load_raw_config),
  with code (pydantic) defaults beneath both; environment overrides still win
  (env > project > global > code). Project config is sparse — brainpalace init writes only values that diverge from what would be inherited; interactive
  prompts default to the global value (flagged "from global") and accepting the
  default leaves the project untouched so it keeps inheriting. New brainpalace config unset <dotpath> removes a project key so it inherits from global/code,
  reporting the new effective source. The dashboard config form shows a
  provenance badge per field and an unset control (with the inherited target)
  backed by POST /config/unset; effective() now returns the inherited
  fallback for project-set keys. --host/--language/--bm25-engine flip to
  inherit-by-default (default=None).
• feat(indexing): per-job embedding token budget guard (indexing.max_embed_tokens_per_job, default 300k). Jobs that would embed more tokens than the budget are blocked with a clear error surfaced to the CLI. force_budget: true on IndexRequest bypasses the guard. limit=0 disables it. Applied in the document pipeline, git-history index service, and session index service.
• feat(folders): record + log folder-add provenance (trigger source).
• feat(estimate): server-less in-process token estimate (brainpalace_server.services.estimate.estimate_tokens_local) so brainpalace init can show the cost estimate before starting a server or writing any index data.

Fixed

• fix(query): reranking no longer returns HTTP 500 for ordinary top_k. With reranking enabled, the stage-1 over-fetch (top_k × RERANKER_TOP_K_MULTIPLIER, capped at RERANKER_MAX_CANDIDATES=100) rebuilt a public QueryRequest, tripping its top_k ≤ 50 validator and surfacing as Server Error (500): Query failed: … top_k Input should be less than or equal to 50 whenever top_k × multiplier > 50 (default multiplier 10 → any top_k ≥ 6). Stage-1 now uses request.model_copy(update={"top_k": stage1_top_k}), bypassing the public-input ceiling for the internal over-fetch while staying bounded by RERANKER_MAX_CANDIDATES.
• docs(cli): brainpalace query --help now documents the --json output schema — per-result keys are text/source/score/chunk_id (not content/file_path), and on failure --json emits {"error": …} (no results key) and exits non-zero, so consumers must check the exit code rather than only the presence of results.
• fix(lifecycle): reap orphan server processes — brainpalace stop --all and brainpalace doctor --reap kill running brainpalace_server.api.main processes that no live registry entry references (leaked servers from a different install surface that hold ports and make auto_port climb). start now also reuses a project's live registry server (find_reusable_server) instead of spawning a duplicate on a climbed port when the project's own runtime.json is missing/stale.
• fix(init): run the pre-index token estimate before writing index data; cancelling at the estimate rolls back the .brainpalace created this run (never an existing one). A pre-existing .brainpalace now prompts delete / keep (resume) / cancel up front instead of silently rebuilding. Index data dirs are created only after the estimate is accepted; the estimate is asked exactly once, up front (removed the duplicate prompt inside the start/watch step).
• fix(estimate): include git-history and session embedding in the pre-index token estimate.
• fix(folders): prune folder records for deleted paths on startup; add brainpalace folders prune.
• fix(git): scope git-history indexing to the project subdir in monorepos (was embedding the entire monorepo history).
• Job "Files" metric showed 0 / 0 for code-only index jobs (most real jobs).
  The 26.6.31 fix threaded real file counts only through the doc-chunker
  callback; the code-chunker progress callback is commented-out/unused, so
  code-only runs never set files_total. Now the real document count is reported
  on the existing 10% "Chunking documents" progress tick (which clears the
  persist-threshold), so files_total is set for every run regardless of
  doc/code mix. Verified live: a watcher re-index now shows e.g. 13 / 13.
• brainpalace update now prints the dashboard URL panel after restarting the
  dashboard (parity with start / dashboard start) — it restarts with
  dashboard start --no-open --json and renders the same render_dashboard_url
  box. Previously the restart was silent about the URL.

---