Deterministic performance counters for CI, one JSON record per run.

Each scenario drives a playground demo through the frame stages the rendering engine runs, in-process with no TTY and no engine GenServer: the theme the engine would resolve (Pipeline.theme/1), view/1, layout preparation, Raxol.UI.Layout.Engine.apply_layout/3, Raxol.UI.Renderer.render_to_cells/2, the border-beam pass (Pipeline.apply_border_beam/2), and the terminal frame (Backends.terminal_frame/2) diffed against the previous one. The stage functions are the engine's own, not copies, and frames/2 is held to the engine's exact output by test/raxol/scoreboard_test.exs, which drives a real Raxol.Core.Runtime.Rendering.Engine through the same keys. Plugin transforms are not applied: they come from the dispatcher's plugin manager at runtime, and the demos run with none loaded, as an app does by default.

It counts three things, and one ratio:

• frame bytes: the size of the ANSI frame payload the engine writes. Exact. A terminal that supports synchronized output also receives the 16-byte \e[?2026h/\e[?2026l wrapper per frame, which is not counted. Every frame is first replayed through Raxol.Terminal.Emulator and must reproduce the screen it was rendered from, character and style (replay_mismatch/1), so the byte count can only fall by sending less for the same screen.
• reductions: the BEAM's unit of work, counted for the measuring process only (Process.info(self(), :reductions)). Unlike wall-clock time it does not move with runner load, which is why pull requests can be compared against master on shared CI runners. It is not perfectly repeatable: runs on one machine agree within a few hundredths of a percent, but two CI runs of the same commit have differed by up to 0.2%, so the reduction metrics carry a 0.5% noise tolerance in priv/scoreboard/metrics.json. Reductions also move with the Elixir/OTP version, which is why the record carries the full toolchain.
• screen memory: the screen buffer the engine holds after the table script, with shared subterms counted once, per cell. Exact. This is the render path's Raxol.Terminal.Cell; the buffer gate's memory rows (merged in with --include) measure Raxol.Terminal.Buffer.
• work ratio: a keystroke's reductions over the first frame's. A renderer that only redoes what a keystroke changed sits far below 1; the table demo is at 0.98, so a keystroke there costs nearly a full redraw. A ratio also falls when the first frame gets slower, so read it next to the reductions it divides.

Every scenario runs once to load code, then three times in fresh processes; the reported reductions are the median. Metric ids, labels, units and the direction of "better" live in priv/scoreboard/metrics.json, which the CI comment and release-note renderer (.github/scripts/scoreboard.mjs) read too. Every metric this module emits must be defined there.

Run via MIX_ENV=test mix raxol.scoreboard.