v0.15.0

374 commits, 393 files since v0.14.0. 153 docs, 88 fix, 88 feat — the project still ships its thinking.

🪶 Surface

• edit_file gains a recovery ladder — whitespace-normalized fallback + escaped-quote decode tier — so over-escaped or re-indented edits self-heal instead of failing.
• symbols overview attaches per-symbol docs inline; single-symbol search shows the code.
• New librarian(action="legibility_scan") — a code-legibility backlog archetype.
• End-to-end usage-friction tracking (friction_target, err_family, overflow_tokens).

🧱 Substrate

• Two-stack retrieval. The daemon-free lite stack (in-process sqlite-vec + one remote OpenAI-compatible endpoint) is now the default build — no Docker, no Qdrant. The full server stack (Qdrant + SPLADE sparse + cross-encoder reranker) is opt-in via --features server-stack. Backend chosen at runtime with CODESCOUT_VECTOR_BACKEND.
• Daemon-free SqliteVecCodeStore + SqliteVecSemanticMemoryStore.
• Windows / EDR support — remote-embeddings path for CrowdStrike VDIs, local-embed-dynamic, Win32 process control, a MinGW+wine build loop.
• peer tool + peer-serve for cross-workspace knowledge; per-session guide-hint ledger.
• codescout-embed split bumped to 0.2.0.

🔬 Evals

• Reconnaissance trigger/output evals and nav/edit harnesses call live MCP tools, with verdicts stored as librarian artifacts. codescout uses codescout to grade codescout.

⚠ Migration

• The default build changed. If you relied on the Qdrant server stack, rebuild with --features server-stack (or cargo rb). Switching backends requires a reindex (vector dimension is backend-specific).