Follow-on to the spawn dedup (#534) and the wait/grid-helper dedup (this PR). Those shared the low-level spawn + the poll-until/grid-query loops. The DEEPER convergence remains: have the e2e runner drive a ttysnap.Harness directly instead of the bespoke Session, and express golden-compare + cast as ttysnap modules (snapshotter/cast_recorder already exist).

Why deferred (not auto-merged): it reworks the e2e golden machinery — the per-scenario grid.txt + grid.sgr compare/update/.actual logic lives in src/test/e2e/{runner,snapshot}.zig, and ttysnap's snapshotter would need to grow that format. That's a deliberate, reviewed refactor of the safety net, not a one-shot auto-merge.

Payoff: running atty's own 34-scenario suite through ttysnap is the strongest proof the framework is production-grade (true dogfooding), and it kills the last of the driver duplication (the two pumpMs + the Session/Harness lifecycle).

Sketch:

• Extend ttysnap's snapshotter to the e2e golden format (txt + sgr, update mode) or add an e2e-specific golden module.
• Make the runner build Harness(.{ golden_module, cast_module }) per scenario; map DSL verbs to Harness methods.
• Keep all scenarios + goldens + per-scenario binary selection + the fragment knob green.