Never scroll the live cursor out of view when anchoring

[dakra]

Problem

On a fresh terminal (cursor near the top, trailing blank grid rows), any window-shrinking minibuffer (vertico/fido M-x, find-file, read-string, …) misbehaved twice:

• While the minibuffer was open, the window scrolled to the blank tail of the grid, hiding the prompt and recent output.
• On exit, the buffer spuriously flipped into copy mode; the "Copy mode" message stomped the echo area (masking e.g. the major-mode guard error from f00d735) and the next keystroke was swallowed by the exit binding.

Root cause: ghostel--adjust-size defers the terminal resize during a minibuffer but still re-anchors. Bottom-aligning point-max in the shrunken window pushed the cursor row above window-start; redisplay then clamped window-point off the cursor — a pure point move no marker can track — and the deferred minibuffer-exit check misread it as user navigation.

Fix

Fix the anchor rather than the check (redisplay's clamp is indistinguishable from a user jump):

• ghostel--anchor-window computes one (START . VSCROLL) from the pixel (GUI) or line-count (TTY) path, clamps START to the live cursor's line, and zeroes the vscroll whenever the cursor sits on the top line. One-directional: a byte-for-byte no-op whenever the cursor is at or below the bottom-aligned start (normal streaming, TUIs, filled scrollback).
• ghostel--window-anchored-p also counts a window as anchored when window-start sits exactly on the (capped) cursor's line, so a clamped window keeps auto-following. Stateless — recomputed from live data on every call, nothing to record or go stale.
• ghostel-semi-char-mode / ghostel-char-mode run goto-char before the forced anchor, so the anchor's point placement (the live cursor) wins and exiting copy mode cannot leave point off the cursor.
• ghostel-maybe-leave-input compares point against the capped cursor position, so a stale cursor beyond point-max cannot fake a jump.

Deliberate navigation (e.g. goto-char after read-string, leaving isearch away from the prompt) still enters copy mode; line mode and the ghostel-inhibit-anchor-functions veto (evil-ghostel) are untouched.

Tests

Five new tests in test/ghostel-modes-test.el (clamp, no-op at bottom, nil/stale cursor, anchored-p recognition, end-to-end no-flip regression), sharing a fixture macro the existing force-anchor test was ported onto.

Verification

• make -j8 all green; evil-ghostel 124/124.
• Live matrix (GUI + terminal Emacs, fido-vertical): M-x text-mode shows the guard error with no copy flip; M-x ignore stays semi-char with the prompt visible and window-vscroll 0 while the minibuffer is open; copy-mode exit lands point on the cursor; filled-scrollback, isearch, and auto-follow paths unchanged.
• Negative control: the same live harness against pre-fix HEAD reproduces the flip in both the pixel and line-count anchor paths.