v1.5.0 — Next-stage agent path

Changed

• Next-stage agent path: init, lite default, Tier-2 loop, HNSW invalidation, search eval

Added

• stele-context init — first-run setup: gitignore-aware index, MCP lite config snippet, doctor next_steps, and recommended agent ritual.
• Doctor guidance — doctor_snapshot now includes next_steps, recommended_mcp_mode (lite), recommended_ritual, mcp_config, token_savings (avoided re-read estimate), and enrichment_preview.
• enrichment_plan — engine + CLI (enrichment-plan) + MCP tool listing hot chunks missing Tier-2 summaries by token mass.
• list_chunk_metadata — lightweight chunk listing (ids, paths, tokens, tier2 flags, optional short preview) for doctor/enrichment without loading full content.
• query suggested_next — sparse/empty composite retrieval returns actionable follow-ups (agent_grep, detect_changes, symbols).
• JS/TS symbol golden fixtures — alias (const X = Y), destructured module.exports, export { A as B }, and exports.Name = patterns covered by regression tests.
• Search quality eval as product — benchmarks/eval_search_quality.py supports --search-mode, --tier2, --tier2-delta, --json, and a recall gate; CI runs keyword + Tier-2 delta; tests in tests/test_eval_search_quality.py.

Changed

• MCP default mode is lite (was standard). Set STELE_MCP_MODE=standard or full for broader surfaces. Lite includes enrichment_plan, bulk_store_summaries, get_search_history, and get_session_read_files so the documented agent ritual works on the default surface.
• Tier-2 invalidation — store_chunk clears semantic_summary / agent_signature when content_hash changes; persist_chunks / sync_chunk_vector refresh HNSW so stale agent embeddings cannot keep ranking.
• Doctor and enrichment_plan use metadata-only chunk queries (not full content for every row).

Fixed

• Eval path mapping on macOS (/var vs /private/var) so corpus recall metrics resolve correctly.
• HNSW/DB consistency after Tier-2 clear on content change (vector replaced with Tier-1).