Emacs backspace fix + buffered terminal output (one flush per frame)

[maxlandon]

Three commits, all GPG-signed. Ideas adapted from / evaluated against the moloch--/readline fork.

Fixes

• fix(keymap) — In emacs mode \C-h was overridden to backward-kill-word; many terminals send ^H for Backspace, so Backspace deleted a whole word. Dropped the override so \C-h falls through to the GNU default backward-delete-char. Regression test asserts the resolved bind.

Performance

• perf(display) — Buffer a whole render frame and flush it in a single write. Refresh() (and AcceptLine/RefreshTransient/ClearHelpers) previously emitted dozens of separate stdout writes per keystroke (hide cursor, moves, clears, prompt, line, highlights, helpers, show cursor); a terminal can show those partial frames as flicker, and each is a syscall. New internal/term buffering layer (BeginBuffer/EndBuffer/FlushBuffer/WriteString/Printf); all render-path fmt.Print sites route through it.
  • Correctness landmine handled: GetCursorPos calls term.FlushBuffer() before the ESC[6n query (unix) / console cursor read (windows), so the position is measured against on-screen output, not buffered bytes.
  • Buffering changes only when bytes are written, never which — the PTY golden-screen tests pass byte-identical. Adds buffer unit tests (coalescing, nesting, mid-frame flush).

Test-only

• test(keymap) — Guards that matchBind's ordering is load-bearing: the default emacs keymap has 15 sequences that ConvertMeta collapses to the same key bytes with different actions, so removing the per-keystroke sort (as the fork did) would make keybind resolution depend on map iteration order. Asserts the collisions exist and that resolution stays deterministic.

Evaluated but NOT adopted

• matchBind sort removal — unsafe (nondeterministic dispatch; see the guard test) and aimed at the wrong cost (ConvertMeta recomputed over all binds per keystroke, not the sort).
• Cursor-position caching — passes the current suite (because ensureInputSpace maintains startRows across scrolls) but has an untested invalidation gap: WatchResize doesn't mark the cache dirty, so a terminal resize would mis-render. Not worth the Incorrect cursor/character rendering in terminal prompt #98-class risk, especially now that buffering cuts the probe's cost.

Verification

go build, go test ./... (incl. display golden-screen + -race), both unix and Windows builds, and golangci-lint run → 0 issues.