Multi-surface parity matrix: renders each fixture session through all four surface projections and pins each one, so "one TEA module renders to terminal, browser, SSH, and MCP" is a checked claim rather than a README sentence.

Raxol.Harness.Surface.Golden pins the raw bytes of ONE surface (the inline/degradation terminal ladder) across three render modes. This module pins ONE render across four SURFACES, and additionally asserts they agree with each other. A regression that teaches the LiveView encoder to drop a wide character, or the ANSI writer to emit SGR before the cursor move, is invisible to a single-surface golden and lands here as both a drift and a parity break.

The shared pivot

Every projection derives from ONE render of ONE fold, so a divergence is always the projection's fault and never the input's:

session (.jsonl)
  -> Raxol.Harness.Projection.project/2   (pure fold, T7)
  -> [Block.t()] -> Block.render/2        (pure view tree)
  -> LayoutEngine.apply_layout/2          (positioned elements)
  -> UIRenderer.render_to_cells/2         (the cell grid)

and from the cell grid:

:structured_json is deliberately taken from the view tree rather than the grid: that IS the MCP surface's input, and pinning it from the grid would test a pipeline nothing runs.

What parity means, per surface

:cells, :liveview_dom, and :ssh_ansi are three encodings of the same grid, so their visible text must be character-for-character identical -- parity/1 asserts exactly that, and it is the check with teeth: each encoder walks the grid independently.

:structured_json comes from the pre-layout tree, so its text is NOT grid-identical (layout wraps, truncates, and pads). Its parity property is containment, checked at the word level: a word the MCP surface reports must be visible on screen. Claiming more would be false precision.

Determinism

Same closure as Raxol.Harness.Surface.Golden's audit, plus:

• fixed geometry (60x24), so layout never reads a live terminal;
• an explicitly constructed theme, so the hash never depends on the global theme registry (which other tests in the suite mutate) -- the same reason Raxol.RATE passes one;
• render_diff/2 is taken against a BLANK buffer, making the ANSI stream a full repaint rather than a function of whatever was on screen before;
• cells are sorted row-major before serialization, so emission order cannot leak into the hash.