Spike: how to avoid re-implementing capture per language

Part of #278

Measured split of the current Python adapter. Not language-specific,
about 1,206 lines: frames 193, screencast 213, snapshot 197, capturer 152, types
115, plus utils, sources, constants and log capture. Genuinely per-language, about
1,595 lines: instrumentation 386, bidi 412, lifecycle 363, transport 188, pytest
plugin 124, backend resolution 122. Four levers to evaluate: move post-processing
behind the wire, push more work into the shared page script, generate contract
types rather than mirroring them, and add a cross-adapter conformance suite.

Acceptance criteria

A decision record naming which levers we take, what moves
where, and what stays irreducibly per-language. Feeds #288 directly.

Size: M.