fix(cli): migrate setup model/provider pickers off simple_term_menu to curses (ESC + Ghostty)

[kshitijk4poor]

What

hermes setup → select a provider → enter the model selection sub-menu → ESC doesn't back out. Same problem affects three sibling pickers. Root cause: these four menus use simple_term_menu.TerminalMenu, whose ESC/arrow-key handling is unreliable across terminals — it conflicts with /dev/tty and, on terminals that emit raw escape sequences (e.g. Ghostty), ESC and arrows misbehave.

This is the same class of bug fixed for the curses menus in #35776 / #35786, but these four sites were never on the curses path.

This PR has two commits

1. fix(cli): migrate the 4 simple_term_menu pickers to curses_radiolist (the hardened helper from #35776, decodes raw CSI/SS3 escapes, consistent ESC):

Picker	Location
Model selection (the reported bug)	auth.py _prompt_model_selection
Reasoning effort	main.py _prompt_reasoning_effort_selection
Named custom-provider model	main.py _model_flow_named_custom
Remove custom provider	main.py _remove_custom_provider

The model picker's pricing column header and unavailable-models block are passed as the radiolist description so they survive the curses clear. simple_term_menu is no longer imported anywhere.

2. refactor(cli): extract a shared curses menu event-loop driver. The three curses menus (curses_checklist / curses_radiolist / curses_single_select) each hand-rolled an identical loop: color-pair init, clear/getmaxyx/refresh, scroll-offset math, the read_menu_key dispatch + NAV_UP/DOWN wrap, flush_stdin, and the fallback. A terminal-behavior change (Ghostty escapes, scroll, a new key) had to be made in three places.

_run_curses_menu now owns that boilerplate. Each menu supplies small callbacks for what genuinely differs: draw_header (returns the item-list start row), draw_row ([✓] vs (●) vs bare), an on_action reducer, an optional draw_footer (the checklist status bar), reserve_bottom, and the numbered fallback. Duplicated event-loop primitives drop 3 → 1. Future Ghostty/terminal work is a one-place edit.

Validation

• Byte-identical refactor proof: a render harness records every addnstr(y, x, clamped-text, attr) call across frames + the return value for 6 cases (checklist, checklist+status, radiolist, radiolist+description, single_select, single_select ESC-cancel). Output diffs clean against origin/main.
• Non-TTY returns the cancel value directly (not the input()-based numbered fallback), matching the old per-menu guard — fixing this also cleared a test-isolation failure where an errant input() polluted later setup tests.
• New test_setup_menu_curses_migration.py asserts each picker routes through curses_radiolist, ESC cancels, and the pricing header is preserved.
• Updated test_terminal_menu_fallbacks / test_reasoning_effort_menu / test_custom_provider_model_switch / test_model_provider_persistence to drive the numbered fallback via a curses_radiolist error (the mock target moved off simple_term_menu).
• 150 menu/setup/browse/plugins tests pass; broader curses/menu/tools/skills sweep 563 pass. Full suite confirmation deferred to CI.

Net production code shrinks the duplicated surface (3→1 event loops); raw line count is roughly flat because the new driver + docstrings offset the removed inline loops — the win is maintainability, not LOC.

Note: #35786 also rewrites these curses_ui.py loops (it added read_menu_key_ex + paging). Whichever merges second will need a rebase on top of the other — I'll handle that.

[teknium1]

Merged via PR #35806 — #35806

Both your commits were cherry-picked onto current main (rebase-merge), so your authorship is preserved per-commit:

• 087be0073 fix(cli): migrate setup model/provider pickers off simple_term_menu to curses
• 8f4c8e7c8 refactor(cli): extract shared curses menu event-loop driver

CI was fully green (all 6 test shards + builds + e2e + lint + nix). Thanks for the clean fix and the byte-identical render-harness proof on the refactor.

Heads up: #35786 also rewrites these curses_ui.py loops — it'll need a rebase on top of this now.