fix(tui): eliminate resume-path terminal corruption in OpenTUI sessions

[usehoplite]

What

Fixes the intermittent terminal corruption when resuming sessions via clai --resume or /history in the OpenTUI UI: blank screen during the session-load window, mouse-event escape sequences printing as garbage, and a broken layout after load (disappearing elements, UI text mixed with transcripts, broken composer).

Root causes (three compounding races)

1. Stale replay overwrote live frames — the runtime host snapshotted the replay buffer before requesting a child repaint and wrote that old snapshot after the reply arrived (up to 5s later), clobbering newer output the child had produced in between. TerminalAttachOutput now gates attachment output: live bytes buffer in order while a repaint request is in flight, the stale snapshot is written only when the repaint is declined, and the buffer is bounded (1 MiB) and disposed with its socket so late replies can never write into a dead terminal.

2. Repaints suspended the renderer mid-attach — the old coordinated flush called OpenTUI suspend(), which leaves the alternate screen, removes the stdin handler, and drops kernel raw mode — exactly while mouse reporting is still enabled, so terminal mouse reports echoed as visible garbage on the blank screen. Repaints now stay inside the renderer: a shape-guarded adapter sets the native one-shot forceFullRepaintRequested latch and lets the existing scheduler emit a complete frame, never surrendering terminal ownership. The bridge repaint handler is registered immediately after the loading screen renders — before any slow resume work — so attaches during the load window get an immediate authoritative frame instead of a timeout + stale replay.

3. Client raw-mode gaps — the durable client restored cooked/echo input mode before writing the terminal reset, and dropped raw mode entirely between /history session switches. Raw mode now spans the entire runtime-client lifetime, a discard pump drains input during connection transitions, and the reset sequence is flushed (bounded error/timeout) before restoring the previous raw state.

Verification

• npm run typecheck — clean.
• test/session-runtime + test/tui-v2 + test/os — 401 tests pass, including new real-PTY integration coverage for delayed-repaint ordering and client handoff reset ordering across detach/exit/switch/signal/auth-failure paths.
• New native renderer regression (real OpenTUI 0.4.5, buffered output) proves a forced repaint rewrites every cell — including blank cells — without suspend/resume, raw-mode drops, or alt-screen churn.
• New PTY continuity test (resize-repaint.pty.py) samples kernel raw mode across 12 repaints with continuous mouse input and includes a negative control proving the detector rejects the old suspend/resume approach. It runs where the native tty writer is available; in sandboxes where OpenTUI's native tty output is inert (verified: frames render with cellsUpdated: 0 even on unmodified code), it skips with an explicit marker. Pixel-level visual verification was not performed; correctness is established through byte-stream, termios, and native-buffer evidence.
• Full suite: 6099 passed; the 17 failures are pre-existing root/sandbox environment issues in unrelated privilege/OS tests (e.g. tests asserting /etc writes fail or sudo prompts — both behave differently as root), none touch the changed code.

Sandbox setup

.hoplite/settings.json now records the effective setup (npm ci + project-local Bun 1.4.2). The system Bun 1.3.1 predates Bun.Terminal, which the session-runtime PTY integration tests require.