Roadmap

The engine grew in phases. Phases 0 through 2.4 have shipped, and the browserproxy remote-browser service that was the original Phase-3 goal is live. What remains is a set of honest levers, not a promise of a full browser. Measured fidelity is on the Fidelity page; the phase-by-phase log with before/after numbers is in the engine's FIDELITY.md and bench/REPORT.md.

Shipped

Phase	What landed
0 — Static renderer	HTML → owned DOM → real CSS cascade (specificity + inheritance) → block/inline flow → AA text/background/image paint → PNG.
1 — Real box model	Floats + clear, flexbox, tables, painted borders, margin collapsing, @media width queries.
1.5 — CSS breadth I	External <link> stylesheet fetch, var() custom properties, anchor click hit-map (Links).
1.7 — Flex + grid	Complete flexbox and CSS grid.
1.8 — Positioning	position: relative / absolute / fixed / sticky.
1.9 — CSS breadth II	Modern colour (rgb()/hsl()), Tailwind variants, border-radius, dark-mode groundwork.
2.0 — Decorations	Linear/radial gradients, background-image: url(), box-shadow, group opacity.
2 — JavaScript	goja bound to a real DOM with fetch()/XHR.
2.1 — SVG	Pure-Go SVG rasterisation (oksvg + rasterx) for <img *.svg>, data: and inline <svg>.
2.2 — Dynamic render	Layout↔JS metric read-back (getBoundingClientRect, offset*, getComputedStyle) + a settle-then-render bounded-fixpoint loop; runtime-injected <script>/<style>/<link>.
2.3 — Selector hack	Sibling combinators ~/+, :checked, :not() — the checkbox-hack that collapses MediaWiki dropdowns; "reduce, don't drop" for unmodelled selectors.
2.4 — JS on hit-map	The RenderWithLinks path shares the JS-settled pipeline, so JS-built links land in the click hit-map.
Remote browser	browserproxy: a WebSocket service that renders server-side with the engine, streams frames + hit-map, forwards clicks/scrolls/keys, and guards against SSRF. Shipping.

Real bold + italic faces replaced the earlier faux-bold synthesis along the way.

Remaining levers (honest, not promised)

• Fidelity residuals — <li> list-style marker discs; icon-font / visually-hidden chrome that renders as text; conic-gradient, CSS filter and mask; SVG <filter>/<mask>/<pattern>/<text>/embedded <image>.
• Performance — large computed pages (pkg.go.dev, go.dev) render slower than Chrome; that is the next big lever, and it is a speed gap, not a fidelity one.
• Interaction depth in JS — more of the CSSOM and event model, so highly interactive SPAs settle closer to their browser state.

These are diminishing-returns work: the mean windowed-SSIM across the five bench pages is already ≈ 0.69, and example.com is at near-parity (0.954).

Non-goals (for now)

• Being a full standards-compliant browser. This is a renderer with an honest, growing feature set — not a Chromium replacement, and not claimed to match Chromium pixel-for-pixel.
• cgo, a bundled browser binary, or a host web view — ever. Pure Go is the whole point.