fix(build): stop build-time projection from caching an empty corpus on Windows

[phil-scott-78]

Problem

On Windows, dotnet run -- build could produce a site whose search index and llms.txt are empty even though every HTML page rendered correctly — and the build still reported success. Every route logged SiteProjection: failed to project /<route>/, skipping with InvalidOperationException: The server has not been started or no web application was configured. Reproduced on a vanilla AddDocSite project; works on Linux CI and in dev/serve mode.

Root cause (instrumented + confirmed)

AuditRunner.StartAsync (a hosted service) fires its initial audit pass fire-and-forget. In build mode that pass resolves LinkAuditor and self-fetches every route through ISiteProjection → RenderedHtmlFetcher → HttpDispatcher.CreateClient(). Startup instrumentation showed this runs before the in-process TestServer has started (its Application is still null), so CreateHandler() throws.

SiteProjection.RenderOneAsync's broad catch treated that infrastructure failure identically to a per-page content error: it returned null, SeedAsync completed with an empty corpus, and AsyncLazy (which only evicts on fault) cached the poison. The search/llms emitters then consumed the empty projection. Linux/dev win the startup race (server starts first); Windows loses it deterministically.

Fix

1. Ordering — AuditRunner runs its initial pass on IHostApplicationLifetime.ApplicationStarted instead of inside StartAsync. That fires only after every hosted service (the web server included) has started, so the self-fetch never races server start. Build-time link-audit coverage is preserved.
2. Fail loud, never cache poison — added SelfFetchUnavailableException; HttpDispatcher.CreateClient throws it for both the un-started-TestServer and no-Kestrel-address cases. RenderOneAsync deliberately does not catch it (catch (Exception ex) when (ex is not SelfFetchUnavailableException)), so an infra failure faults SeedAsync, AsyncLazy evicts, and the next access retries instead of baking an empty corpus.

The report also suggested an "abort the build if 0-of-N routes project" guardrail; intentionally skipped to keep the change surgical — fixes #1 + #2 already make this failure mode either work or fail loud.

Verification

• Full solution builds clean; 880 unit + 48 integration tests pass (integration suite exercises the real TestServer self-fetch path).
• DocSiteScaffoldExample build: search index n=2 with both docs, llms.txt lists both pages — 3/3 runs (previously a deterministic empty build on Windows).
• BareHostSearchExample (~100-doc corpus): 85 pages, 484 heading-level search docs.

Tests added/updated

• HttpDispatcherTests (new) — un-started TestServer and no-address server both surface SelfFetchUnavailableException.
• SiteProjectionTests — infra failure during seed faults rather than caching empty, then retries to a populated corpus once the server is up.
• AuditRunnerTests — updated for the ApplicationStarted gate.