fix(agent): width-aware interactive prompt/status line (#209)

[justrach]

Problem

Agent.prompt emitted the full model/effort/provider/mode/cwd/context status line with no width budget, so a narrow terminal pane soft-wrapped it mid-badge (e.g. splitting codex) and left the cursor visually inside a label:

[gpt-5.6-sol · High · co

readline.zig only reacts afterward by moving the input to a fresh row; by then the status string has already wrapped.

Fix

prompt() budgets each segment against termCols():

• The model plus the settings that disambiguate the cursor (effort, provider, mode badges) are offered first; cwd/context/cache/cost are dropped when they no longer fit.
• The context meter outranks cwd for the budget (it is the urgent, changing signal near the compaction threshold) but still renders after cwd, so the familiar left-to-right order is kept when both fit.
• The line reserves the fixed frame ([ … ] › ) plus a column of slack, so nothing but a model name longer than the pane can wrap; readline then gives the input its own row as before.

Progressive disclosure at increasing widths (captured from a real PTY):

cols=28   [deepseek-v4-pro] ›
cols=34   [deepseek-v4-pro · Medium] ›
cols=50   [deepseek-v4-pro · Medium · codegraff] ›
cols=120  [deepseek-v4-pro · Medium · codegraff · cwd /…] ›

The wcwidth-style dispWidth/codepointWidth table moves from repl.zig into term.zig (the shared terminal-utilities leaf) so the TUI table math and this budgeting share one definition; repl.zig delegates and keeps its #142 test.

Tests

• New scripts/test-pty-narrow-prompt.py — drives a real 28-col PTY and asserts the line fits, badges stay whole, and cwd is dropped, while a wide pane still shows cwd.
• scripts/test-pty-repl.py widened to 200 cols so its badge-state session fits the full line.
• zig build test 184/184, zig fmt clean, full PTY suite green.

Closes #209