A composer: F1 opens a full editor for writing a post - #25

The command line is a command line: it grows and wraps, but ⏎ sends, so
anything longer than a couple of sentences is written somewhere else and
pasted in. F1 opens a full-screen editor instead — undo, find, selection,
mouse and a caret over wrapped rows — and sends the result as one command.

This is the one place `MultilineEditControl` is right. It is ruled out for
the command line precisely because ⏎ there has to send rather than insert;
a composer is the opposite case, so none of that behaviour is written again.

What is sent (`ComposeMessage`, Core): the buffer is the whole command,
verb and all, with its line breaks joined by `%r` — a MUSH stores a post
as one string and renders the breaks itself, which is what `+bbpost` and
`@mail` bodies want. Blank rows at the ends are where the caret was left
and are dropped; interior ones are paragraph breaks and survive as `%r%r`.

⌥L switches escaping. In `literal` the body's `% [ ] { } ;` and `\` are
escaped so the post shows what was typed; in `as typed` nothing is. The
escaping runs per line *before* the breaks are joined in, so the `%r` this
writes is never itself escaped — the other order produces `%%r`, which
posts the characters "%r" instead of a line break on every line.

Modal, which is what makes it possible: `PinFocusToArmedBar` would fight an
editor needing real focus for ever, and it stands down while the main window
is inactive. It refuses over an open settings screen — that overlay takes
paste off the driver because its screens have no focusable target, and this
one is exactly such a target, so both would fire on a single paste.

Drafts are per character, in memory for the life of the run. Keyed by the
window's owner, never `_active`. Closing raises the event that stores one,
so sending closes first and forgets after.

F1 is claimed in MacroKeys.AppShortcuts and answered before the settings
screen lookup, so a future unclaimed F-key cannot silently open this. That
takes F1 off MacroKeys.Bindable, which moved one macro test to F12.

New views `compose` and `compose-literal`. Build clean, 0 warnings; Core
864, Tui 1542, Graphics 83, Scripting 42, Web 37 — all passing.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015nuKnWthnELNkrd86q5KWN

`AnyOverlayOpen` is this app's statement that a modal surface owns the
screen, and the composer was missing from it — so ⌃R and the other guards
that read that flag acted while a post was being written. It is in the
list now.

F1 correspondingly refuses over *any* surface rather than only over a
settings screen. That guard was written for the paste hazard specifically
(SettingsOverlay listens at the driver because its screens have no
focusable target, and the editor is one), but the rule it was enforcing was
always the general one. The refusal names the surface in the way rather
than saying "a surface", like every other refusal here.

Not changed: the palette, quit prompt, message log and settings screens
still open over each other, which they did before this branch. Making them
mutually exclusive is a behaviour change to five existing surfaces with its
own test surface, and ⌃P has to keep closing the palette it opened.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015nuKnWthnELNkrd86q5KWN

The guard was one-sided: the composer refused to open over a surface, and
every surface still opened over the composer, because a global shortcut
runs before any window sees the key. ⌃P, F2–F9 and ⌃P ▸ Show client
messages now refuse while a post is being written, naming what is in the
way and saying that Esc keeps the post.

⌃Q is deliberately not one of them. A client you cannot leave from a modal
is worse than one that stacks a prompt, so the quit prompt still opens —
and QuitFactsNow counts the unsent post instead, which is the answer that
helps: it says what leaving would cost. The composer keeps drafts in memory
only, so quitting is precisely the thing that loses them. The open window's
buffer is counted separately from the stored drafts because it is not in
that dictionary until it closes.

Still not changed: the pre-existing surfaces open over each other as they
did before this branch.

Core 864, Tui 1546, Graphics 83, Scripting 42, Web 37 — all passing, build
clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015nuKnWthnELNkrd86q5KWN

Both conflicts were "keep both". `--view compose`/`compose-literal` and
`--view links` are separate arms of the same snapshot switch; the view list
and CLAUDE.md each gain both sections rather than one replacing the other.

Also fixes two review items on the way past:

- `OpenOverlayName`'s doc comment had been stranded above
  `ComposerIsInTheWay` when that method was inserted, so it documented the
  wrong member and the other had none.

- ⌃Q counted a reopened post twice. After Esc and F1 the same text is in
  the draft store *and* in the open window — the store kept it on the way
  out and the window is holding it again — so the prompt said "2 unsent
  drafts" for one post. The open composer's own character is now excluded
  from the stored count.

Core 883, Tui 1601, Graphics 83, Scripting 42, Web 37 — all passing, build
clean. compose, links and tint-input all still render.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015nuKnWthnELNkrd86q5KWN