fix(frontend): recover clipped Codex terminal via onRender convergence re-fit

[harshitsinghbhandari]

Problem

Closes #280. The Codex terminal renders only in the top ~half of the pane; the cursor sits mid-pane, everything below is empty bg-terminal black, and the user can't type or scroll.

Root cause (per the triage RCA on #280)

FitAddon derives the row count by dividing the pane height by the renderer's measured cell box. That box is measured asynchronously: the WebGL renderer loads after term.open(), and the monospace font's real metrics resolve a frame or more later. The early fits (rAF + two timeouts + fonts.ready) can all run while the cell box is still too tall, so xterm under-counts rows and sends that short grid to zellij over the mux. The grid then freezes: the only remaining fit trigger is a ResizeObserver on the host, and the host's height:100% box never changes when only the row count is wrong, so the bad grid sticks for the session's lifetime.

Fix

Add a recovery re-fit that is not gated on the host box changing size:

• onRender convergence loop — each renderer repaint re-proposes dimensions from the current measured cell box and re-fits when they differ, converging the grid to the true row count once the cell metrics settle. proposeDimensions() returns undefined until the cell box is non-zero, so a fit is never accepted from an unmeasured cell. The listener detaches once the proposal holds for a few frames (or a hard re-fit cap is hit), so steady-state content renders cost nothing.
• window resize listener — OS-window and monitor/DPR changes move the true cell box without touching the host's height:100% box, so the existing ResizeObserver misses them; observe window directly as a session-long recovery path.

Both are torn down on unmount. The non-integer lineHeight: 1.35 and the post-open renderer load order were the suspected aggravators of the over-measure; this fix is renderer- and lineHeight-agnostic because it converges on whatever cell box the renderer ultimately settles on, rather than trusting the first measurement.

Verification

• npm run typecheck clean.
• npm test — 140/140 pass.
• The terminal-fills-the-pane behavior depends on the live WebGL/canvas renderer measuring real cell metrics, which jsdom can't exercise, so it isn't covered by an automated test. Recommend a manual smoke check (open a Codex session, confirm the grid fills the full pane and survives an OS window resize) before merge.