v5.7.0

Changed

• /vkm-design stops ordering from its own menu (ADR-0090). Surface-tier generate now
  gates on a live dossier (three URLs for this brief, stolen technique + spacing
  attitude) before any direction; lineages are grammar not costumes; contemporary.md /
  new craft-tech.md decode what you saw on those pages instead of offering a 14-move
  picker. That was the field failure: every agent (cloud, local, Cursor, keyboard) restyled
  Inter/indigo into Fraunces-on-warm-paper or dark-luxe gold and called it variety.
  slop-check.mjs prices the second defaults too (Geist/Roboto, Tailwind #3b82f6).
• /vkm-ui-judge measures space, not just contrast. layout-checks.mjs flags
  viewport-edge flush, tight card padding, content overlap, component drift and cramped
  stacks; ui-audit.mjs merges them into report.json and writes geometry.json. The
  visual loop and the design-ui acceptance gate treat missing gutters as defects.
• Cursor pays the memory protocol once (ADR-0089). --rules with both agents and
  cursor writes the full block only to AGENTS.md; ~/.cursor/rules/obsidian-memory.mdc
  and the project .mdc become a short pointer (reinstall replaces a legacy full copy).
  Core arbitration: User Rules must not dump START_HERE / PROJECTS / SESSION_LOG
  whole. Startup is hybrid/assemble_context, not a ritual START_HERE.md read. Install
  docs tell you not to paste the block into Cursor User Rules. This repo's
  00-stack.mdc is alwaysApply: false (the stack text already lives in AGENTS.md).

Added

• DeepSeek V4 as a first-class research backend (API, OpenAI-compatible). The
  obscura-web research stack no longer hard-requires a local Ollama: the curation /
  query-expansion / translation / lead / summarization calls now route through an
  OpenAI-compatible /chat/completions client selected by OBSCURA_LLM_PROVIDER
  (auto = DeepSeek when DEEPSEEK_API_KEY is set, else Ollama; explicit
  deepseek/openai/ollama/none). The API key is read from DEEPSEEK_API_KEY
  only — never written to disk, git, mcp.json, or any log, and redacted from every
  error the module surfaces. OBSCURA_LLM_PROVIDER=ollama (or unsetting the key) returns
  every call to local models when the API gets expensive.
• DeepSeek model detection + tuning: the effort gate (guard-effort-gate.mjs)
  recognizes deepseek-* and never proposes a Claude /model to it; the SessionStart
  vault-context hook injects a short DeepSeek-only block (thinking ON by default, reasoning
  in reasoning_content, reasoning_effort flash/pro split, key-env reminder); and the
  en/es/examples agent-profiles.md rows plus new AGENTS/deepseek-v4.md notes document V4.
• --research-llm <auto|deepseek|ollama|none> and --deepseek-model <model> installer
  flags, wired to OBSCURA_LLM_PROVIDER / OBSCURA_DEEPSEEK_MODEL on the obscura-web MCP
  server (the non-secret half of the lever — the key itself is never persisted).

Fixed

• obscura_research dying by MCP-transport timeout: the SearXNG-less single-page scrape
  fallback is now raced against the call's own deadline, and a plain-HTTP GET fallback
  (nativeHttpGet) serves DuckDuckGo HTML / Bing RSS when the obscura binary is missing or
  hung, so a dead headless browser returns scanned:0 instead of blowing the 60s wall.