feat: declarative TUI end-to-end test harness (testty-style)

## Idea: declarative TUI end-to-end test harness (testty-style)

### Current state

graith already has good test coverage at two layers: a **daemon integration harness** (`internal/integration/integration_test.go`, `//go:build integration`, drives `SessionManager` + `Server` over a socket) and **PTY-level unit tests** (`internal/pty/*_test.go`, including render and scrollback tests).

### Gap

There's no higher-level **declarative TUI end-to-end** harness that drives the real `gr` TUI through a PTY. agentty ships exactly this ("testty"): YAML scenarios, predicate-driven waits ("eventually the screen contains X"), frame/screen snapshot assertions, and shareable **proof artifacts** (text/HTML/GIF, content-hash cached).

### Proposal

A declarative TUI E2E harness for graith's interactive surfaces — the session picker/overlay, the approval overlay, the dashboard, and the attach/passthrough flow — that today only have unit coverage. Scenarios spawn `gr`, send keystrokes, and assert on rendered frames, producing reviewable artifacts in CI.

### Notes / open questions

- This is dev-infra, not a user feature — but the interactive surfaces are exactly where regressions hide.
- Build on the existing `internal/pty` render/vterm machinery for frame capture.
- Scope: assertion model (substring vs full-frame snapshot), artifact format, flake-resistance (predicate waits, not fixed sleeps).

_Source: agentty `testty` PTY-driven E2E framework (competitive analysis)._

Start by reading internal/integration/integration_test.go and the render and scrollback tests under internal/pty to understand the existing harnesses and frame machinery. Then inspect how the gr TUI exposes the session picker, approval overlay, dashboard, and attach/passthrough flow. Done means declarative scenarios can drive gr through a PTY, wait on predicates, assert frames or snapshots, and produce reviewable CI artifacts.

Written by the indexing model from the issue text.