docs: address spec review feedback on build performance design

docs/superpowers/specs/2026-03-19-eleventy-build-performance-design.md

@@ -39,6 +39,8 @@ DEBUG=Eleventy:Benchmark* npm run build 2>&1 | grep -E "Benchmark|og|unfurl"
 
 This captures both Eleventy's internal timings (collections, transforms, filters, rendering) and the two before-hook timings in a single build run.
 
+**Note on OG subprocess output:** The OG hook uses `execFileSync` with `stdio: "inherit"`, which means the subprocess writes directly to the parent's stdout/stderr — its output bypasses any pipe or grep on the parent process. The `console.time("[og] image generation")` and `console.timeEnd(...)` calls on the parent process measure the full wall time of the entire subprocess-spawn loop and will appear in filtered output. Individual per-batch log lines from the subprocess itself (`[og] Generating...`) will appear unfiltered inline.
+
 ---
 
 ### 2. `hash` filter memoization
@@ -47,7 +49,7 @@ This captures both Eleventy's internal timings (collections, transforms, filters
 
 **Problem:** `readFileSync()` + MD5 on every call. With 500+ pages each calling `| hash` for 2–3 asset paths, this is ~1000+ redundant disk reads per build.
 
-**Fix:** Wrap with a `Map` cache, cleared on `eleventy.before`:
+**Fix:** Wrap with a `Map` cache, cleared on `eleventy.before`. Declare `const _hashCache = new Map()` inside the config export function body, immediately before the `addFilter("hash", ...)` call — matching the placement of `_dateDisplayCache` at line 550 (not at module scope):
 
 ```js
 const _hashCache = new Map();
@@ -123,6 +125,7 @@ eleventyConfig.on("eleventy.before", async () => {
 - Manifest stored in `.cache/` — same gitignore and passthrough-copy rules as existing unfurl cache.
 - `writeFileSync` is used (not async) to guarantee the manifest is written even if the process exits unexpectedly mid-batch. On failure, next build falls back to full prefetch.
 - The content walk still runs on every build (cheap: local FS reads of small frontmatter). Only the prefetch calls are skipped.
+- Writing `[...urls]` (not `[...seen, ...newUrls]`) is intentional: it prunes URLs for soft-deleted posts on the next prefetch-triggering build, preventing unbounded manifest growth.
 
 ---
 
@@ -141,5 +144,5 @@ No new files. No new dependencies.
 1. `console.time` output visible in production build logs.
 2. `DEBUG=Eleventy:Benchmark*` output available for baseline comparison.
 3. On a build with no new posts: `[unfurl] No new URLs — skipping prefetch` logged.
-4. Hash filter: same cache-busting behaviour as before (values change when files change between builds).
+4. Hash filter: modify a referenced asset (e.g. `css/style.css`) between two builds and confirm the `?v=` query string changes in rendered HTML. No log output expected — verify via view-source or build diff.
 5. No change to rendered output.