feat(linkedin-studio): research-engine config layer — sources + scoring modes + MCP profile (§5 slice 2a) [skip-docs]

Declaration/config groundwork that slice 2b's trend-spotter upgrade reads.
Standalone (no agent wiring yet — that's 2b), mirroring slice 1's pattern;
[skip-docs] for the same reason slice 1 was — user-facing docs land when 2b
wires the engine live.

- references/trend-scoring-modes.md: methodology SSOT for two rubrics —
  kortform (feed post, timing 20%) + long-form (chronicle, depth 25% / timing 10%,
  per tema-research-motor-spec §4.2). Both sum to 100%. trend-spotter renders from
  this in 2b instead of inlining a matrix (S12-consistent).
- config/trends-sources.template.md: shipped generic source-list defaults →
  user override at ${LINKEDIN_STUDIO_DATA}/trends/sources.md (data-dir, survives
  reinstall; same template->data-dir pattern as user-profile).
- user-profile.template.md: new "Research Tooling" section — declared research MCPs
  (Tavily/Gemini/Perplexity/Other) + WebSearch/WebFetch floor. 2b routes MCP-first.
- setup.md Step 3f + onboarding.md Phase 2: ask "which research MCPs?" -> profile.
  Store only what the user declares; no hard-coded MCP names.
- test-runner.sh: EXPECT_REFS 26->27; generalized the M0 +1 delta-guard into a
  named-post-M0-additions guard (POSTM0_REFS) so a legit later ref doc passes while
  the anti-masking intent holds. Gate green 84/0/0.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RBMKqPSVbvSZHtQ4heM1UY

commands/onboarding.md

@@ -154,6 +154,7 @@ file must contain no `<!-- VOICE_PLACEHOLDER -->`.
 3. Job title / role
 4. 5 expertise areas (these become your content pillars)
 5. Target audience description
+6. Research MCPs connected (Tavily / Gemini deep research / Perplexity — or "none"; WebSearch + WebFetch are the always-available floor). Store only what they name — don't invent MCP names.
 
 Save to `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/profile/user-profile.md`.