One join point, two audiences: why the MCP server reads the build artifact, not the source

Pre-joining costs bytes, and the invoice is legible. The index the agent reads is roughly 1.4 times the size of the 1,312,577 bytes of hand-edited JSON it comes from ^[5][1], which is the reverse of what an API layer would have shipped. It is also only about 11 percent of the 16,644,215 bytes the build writes in total ^[6][2]. The deployed endpoint is mostly freight: 21,508 bytes of hand-written JSON-RPC over POST, no MCP SDK ^[11], carrying something like 88 times its own weight in data it did not compute ^[7].

The reason for the freight sits in api/mcp.mjs, whose first line of real work destructures the imported index instead of deriving anything ^[10]. Put an API in front of the data and that file has to know how a name becomes a slug, which parents count as ancestors, and which adoption edges are shown, all of which the page renderer already knows ^[9]. Slugs, lineage and adoption edges are joined in exactly one place, so the crawler and the model cannot disagree ^[4].

The two indexes are sized by different tolerances. A crawler collects context by walking links: the build reports 96,843 internal links across 1,245 pages ^[7], about 78 per page ^[4], and there are 65 more pages than there are entities ^[3]. The human clicking through is the retrieval system ^[14]. An agent has no comparable budget. A stub answer buys either a guess or four more tool calls ^[15], so the record ships the credited origin game and its developer, the prose essay, parents, children, everything downstream and the later adopters, already joined ^[16].

The overview tool spends its description on definitions rather than schema: origin means the first notable shipped implementation rather than invention, which is why over-the-shoulder aim is credited to kill.switch in 2003 and not Resident Evil 4 in 2005 ^[18]. The same tool states that 949 of 1,180 entities carry a verified Wikipedia permalink and 231 carry none ^[19], roughly a fifth of the base with no citation to offer ^[5].

What reading an artifact actually costs is freshness, and the build time keeps that small: under half a second with zero npm dependencies ^[1], timed at 0.40s and 0.39s on repeat runs ^[20], and a deleted output directory comes back inside the same window ^[8]. Because mcp-index.json is committed rather than ignored, a claimed rebuild is checkable in a diff ^[20]. The residual risk is narrower than staleness in general. It is a deploy in which the pages and the index came out of different runs, and one join point does not close that seam on its own.