feat(fixtures): consolidate corpus around a flagship demo fork + mutation tests

[arielshulman29]

What

Consolidates the e2e fixture corpus from 21 miniature apps to 11 meaningful ones, centered on a flagship fixture forked from the ora-air demo, and adds a multi-build mutation-test harness for the build-N+1 bugs the one-shot golden layer can't see. Three product bugs the new corpus immediately surfaced are fixed in the same PR.

Fixture corpus (21 → 11)

Fixture	Role
flagship (new)	Fork of the ora-air demo, enriched: dynamic segment (destinations/[slug] + generateStaticParams), hand-owned llms.txt route, /guide via @next/mdx, sitemap.ts, hand-authored /destinations twin, gated /account page, two mcp-handler servers (public + withMcpAuth-gated), withAx(manifest, botGate) middleware. Absorbs middleware, mcp-adapter, mcp-adapter-gated, mcp-multi-server, llms-txt, openapi, discovery, config-overrides.
flagship-pages (new)	Full Pages Router port of the flagship IA — the one deep axis that forks the router model and everything downstream. Absorbs pages-bare, pages-mcp, pages-webmcp-declarative.
webmcp (new)	Declarative <form toolname> + imperative registerTool() on one page (merges webmcp-declarative + webmcp-imperative, keeps the false-positive-prose regression).
next-auth (new)	Thin single-axis fixture: real next-auth dependency + [...nextauth] mount — net-new real-build coverage for provider detection.
bare, bare-js, deploy-variants, edge-cases, monorepo, markdown-twins, hybrid	Kept unchanged (absence baseline / JS axis / basePath / WebMCP decoys / nested app / twin-ladder rungs / app+pages union).

Design notes:

• Zero secrets: the demo's Clerk wiring is replaced with a stub withMcpAuth verifier + hand-written static RFC 9728/8414 .well-known documents — same detection surface, deterministic CI builds. Provider detection stays covered by unit tests + the next-auth fixture.
• The golden/conformance pipeline needed no changes — snapshot discovery is dynamic. Only fixtures-integration.test.ts (every deleted fixture's assertion mapped to an explicit new home) and dogfood-middleware.mjs (retargeted to flagship, plus a new composed-middleware probe: scripted UAs still get bot-gated through withAx's second argument) reference fixtures by name.
• The real win is CI wall-clock (17 → 8 report builds) and maintenance surface; the flagship additionally tests detector composition, which no isolated fixture could.

Mutation-test harness (pnpm mutation:test, wired into CI)

Copies the flagship's tracked sources to a tmp dir (one absolute symlink reuses its pnpm node_modules), then runs build → mutate → rebuild sequences with targeted delta assertions:

1. Idempotence — steady-state rebuilds are byte-identical (report + every twin), robots.txt append never doubles, nothing is deleted on a no-change rebuild.
2. Auth-method change — flipping the declared entry auth api_key ↔ oauth2 propagates to the report, catalog entry, and auth.md with no remnant of the old method, and never re-opens the MCP review gate (a declared auth marks its mount reviewed).
3. User-edit preservation — a filled-in scaffold TODO survives the next build.
4. Staleness — deleting a page removes its orphaned twin and manifest entry.
5. Additive — a new page and a new MCP tool land in twins, cards, and the report.
6. Manifest drift — removing the gated page refreshes gatedPaths via the prebuild.

Bugs the new corpus caught (fixed here)

• MDX pages were invisible to the router model: /guide built and got a twin, but was missing from the serving manifest — agents requesting it got the 404 wayfinding page for a route that exists. buildRouterModel now takes pageExtensions (from next.config) and counts page.mdx/.mdx pages as routes exactly when Next serves them, staying conservative otherwise.
• 404.md was swept as a stale twin on every build after the first: it carries the generated-by marker but is lifecycle-owned by agent-404, so the sweep deleted it and its owner rewrote it — a phantom markdownTwins.deleted entry CI never saw (it always builds once from clean). The sweep now skips auth.md/404.md; stale-twin deletions are also logged per-path like server cards.
• The init multiselect garbled on arrow keys: the raw-mode selector paused the shared readline interface but resumed stdin, leaving readline's keypress listener live — arrow-up recalled the previous question's answer from history and painted it over the option rows; printable keys were echoed at the cursor. Reproduced under a real PTY, fixed by suspending readline's keypress listeners for the duration of the selector (Node's keypress machinery re-attaches on restore — verified with follow-up questions). Also guards a PTY reporting columns: 0, which truncated every row to the empty string.

Verification

• 736 unit tests, 8/8 report/twin/card snapshots (hermetic reports:regen), 8/8 ARD conformance, 18 dogfood probes, 29 mutation probes, typecheck, prettier — all green.
• Only golden diff from the fixes: flagship's 404.md grew one line (/guide now listed as a real route).

🤖 Generated with Claude Code