# Upstream interface capture `scripts/aphrody/site-mirror/mirror.ts` invokes the native Aphrody Web owner. Capture runs on a qualified SSH host in a separate state directory. It never changes a source checkout or the active production release. ```sh bun scripts/aphrody/site-mirror/mirror.ts plan /path/to/new-capture bun scripts/aphrody/site-mirror/mirror.ts capture /path/to/new-capture bun scripts/aphrody/site-mirror/mirror.ts build /path/to/capture /path/to/new-preview ``` The native mirror follows bun.com and www.bun.com links, up to 2,500 pages, 12,000 assets and depth 20, with four concurrent requests and a 300 ms start delay. CDN traversal is disabled. These are limits, not measured coverage. The original HTML, CSS, JavaScript and asset bytes remain in `upstream/`. The native JSON report and error log remain beside it. For sitemap and robots coverage, use the separate native persistent frontier: ```sh aphrody web crawl https://bun.com --sqlite /path/to/capture/frontier.sqlite \ --max-depth 20 --max-requests 2500 --concurrency 2 \ --allowed-domains bun.com,www.bun.com --per-host-delay 300 \ --max-per-host 2 --max-retries 2 --obey-robots --sitemaps --json ``` This frontier can be reopened. Compare its discovered URLs with the mirror report before claiming a complete capture. Linked assets and dynamic JavaScript imports need separate inspection; the collector does not imply browser parity. Public assets are retained in private capture state, not committed to Git. The reference sitemap can contain many pages absent from ordinary navigation. `complete-reference.ts ` plans missing canonical bun.com reference URLs from the frontier SQLite database. `--apply` acquires their original HTML through `aphrody-web response`, with two workers, a 300 ms request interval, a 2 MB response limit and at most 20,000 URLs with three attempts each. HTTP 429 uses the actual `Retry-After` header, or exponential backoff when absent. A wait above five minutes defers the batch rather than retrying before the server asks. Existing captured bytes are retained. SHA-256 receipts in `reference-capture.jsonl` allow interrupted runs to resume. Truncated responses and HTTP errors remain failures. The completion report explicitly separates this reference family from other unlinked routes. Rebuild a new preview after the collector ends; keep the previous preview. Native mirroring can rewrite an old documentation link to an alias while storing only the HTTP redirect target. `repair-redirects.ts ` plans these missing aliases; `--apply` observes each redirect with the native response owner and writes an alias only when its confirmed canonical target is already captured. No redirect destination is guessed. Redirect evidence is stored beside coverage. The preview retains captured DOM order, layout, JavaScript and navigation. An additional stylesheet and script map upstream color variables onto the current Aphrody M3 semantic roles and apply shape tokens. Captured typography retains upstream `--font-display` and `--font-sans`, mapped to the M3 brand and plain typeface roles. The captured stylesheet declares Archivo for display and its existing system stack for plain text. Bun OS surfaces use Aphrody Google Sans Flex independently, without changing captured line breaks. Bun OS navigation exposes release downloads, docs, guides, plugin skills, the visual code graph and measured benchmarks. Installer and canonical URLs change to bun.aphrody.app. Upstream analytics are removed. Upstream benchmark values retain their original provenance and measurement scope. `mirror-coverage.json` records each original and adapted SHA-256, missing local HTML references, the exact release source SHA and collection limits. Missing references and unmeasured screenshot or interaction parity remain visible. `aphrody web browser snapshot` exposes readable snapshots. The native owner `aphrody-web launch --host 127.0.0.1 --port 9298 --workers 1` exposes Obscura CDP, including actual `Page.captureScreenshot`. `capture-browser.ts` records PNG bytes and hashes, the viewport, source DOM links/buttons/scripts, and actual installer and benchmark tab smoke observations. For a trusted local preview, scope `OBSCURA_ALLOW_PRIVATE_NETWORK=1` to this isolated browser process only. `serve-preview.ts` binds 127.0.0.1:8799 and applies the production CSP; executable inline scripts become deterministic same-origin assets, preserving source order and module attributes. JSON data remains inert. No CSP relaxation or global network configuration is required. Screenshots establish capture evidence, not pixel equality. M3 colors and shapes are intentional visual differences; compare geometry and behavior separately. The installed Obscura `Emulation.setEmulatedMedia` currently applies reduced motion but not color scheme. A requested dark viewport is not evidence of a dark rendering: inspect `observedTheme` and qualify that view separately. The 2026-10-11 native capture reported 3,178 requests, 2,196 pages, 966 assets and 694,723,122 bytes, with 16 failed resources. The sitemap frontier still contained 16,214 pending URLs; the separate reference collector continues resumably. These counts do not establish complete coverage. At 390 by 844, upstream and preview both selected the five tested installer and benchmark tabs. An isolated page clipboard collector observed the upstream Windows installer and its Bun OS URL replacement. Neither rendering exposed a theme button or a visible search input after the search action. Both reported the mobile menu's `aria-expanded` as false after its opening action. Menu and search parity therefore remain unqualified. The initial M3 font override reduced the heading height by 37 pixels and moved subsequent tabs upward by 63 pixels. After retaining captured font metrics, all six observed boxes matched upstream in x, y, width and height at both 390 by 844 and 1280 by 900. The five tested tabs remained selectable. These twelve box observations do not establish pixel equality or cover other routes. The separate `brand-geometry-receipt.json` records source SHA, adaptation CSS SHA, generator SHA and screenshot hashes; the original coverage manifest predates this staging-only CSS change. Stage the result under an upstream preview route. Before promoting its home page, compare light/dark desktop and narrow layouts, navigation, search, installer tabs and copy actions, benchmark tabs, workflow controls, terminal replay, release links, responsive menus and keyboard behavior. Keep the release gateway authoritative for `/docs`, `/guides`, `/knowledge`, `/plugin`, `/graph`, `/benchmarks`, `/downloads`, installers and runtime version metadata. Preserve the previous static release and its Nginx configuration for rollback. Sources inspected: [Bun](https://bun.com/), the installed `aphrody web mirror --help`, `aphrody web crawl --help` and `aphrody web browser snapshot --help`. No screenshot parity or complete site coverage is implied by these documents.