Stage 3 partial: LinkedIn MCP server (OAuth, 9 tools, kill-switch, refresh lock)

What ships (testable without live LinkedIn, 27 new tests):
- apps/mcp-linkedin/src/oauth.ts: auth URL builder + HMAC-signed state validation (CSRF + tamper + expiry)
- apps/mcp-linkedin/src/refresh-lock.ts: advisory-lock helper for token rotation (Plan TEA gap 3); concurrency test verifies 4 attempts → 1 succeeds + 3 denied
- apps/mcp-linkedin/src/kill-switch.ts: 30s-cached feature-flag query (Plan Objective 8 + TEA gap 10)
- apps/mcp-linkedin/src/tools.ts: 9 Zod tool schemas matching Plan §3.2 (whoami, auth_status, create_post, create_article, upload_media, create_post_with_media, delete_post, get_post_metrics, get_profile_stats)
- apps/mcp-linkedin/src/server.ts: validateToolCall + outletForAuthorUrn pure helpers

What defers to live-LinkedIn session (gate 0.9):
- 3.1 OAuth round-trip with real auth URL → callback → token row
- 3.4-3.7 Live throwaway test posts + delete-within-5min audit
- 3.9 Fail-safe halt with Telegram webhook
- 3.12 MCP stdio transport wired to @modelcontextprotocol/sdk

106/106 tests pass across all packages and apps.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

apps/mcp-linkedin/package.json

@@ -11,7 +11,6 @@
     "lint": "echo 'lint pending'"
   },
   "dependencies": {
-    "@modelcontextprotocol/sdk": "1.0.3",
     "@stargue/linkedin-client": "workspace:*",
     "@stargue/schema": "workspace:*",
     "@stargue/observability": "workspace:*",