macos: one tagline rule for both hosts, a kqueue beside the inotify, and effects that compile

Three things this shell had its own copy of, and in each case the fix is that
it stops having one.

**The tagline band.** A pane tag draws at `gui_tagline_font_percent` of the
body face and the band it sits on shrinks with it, while the grid row stays
body-sized — so something has to decide where the shorter band sits in the
taller row. This shell decided by centring, always, which is precisely the case
`config.gui_topbar_pane_border_px` exists to prevent: the topbar's unused
half-band meets the first pane tag's unused half-band and the window background
shows through the seam. The strip is as wide as the bands are short — on a
20-pixel cell, 4 physical pixels at the default 82%, 10 at 50%, 14 at 30% — so
it grew as the tagline face shrank and read as "the tagline is wrong on the mac"
rather than as one missing rule. The rule is `pardes.taglineBandOffset` in the
core now and both pixel hosts call it: row zero bottom-aligned, the first
pane-tag row top-aligned, the two joined by `gui_topbar_pane_border_px` in the
theme's scrollbar-track colour, every row between centred, and a `Tagbottom`
band on the final row flush with the window edge — with the sub-cell strip
beneath it painted in that band's own colour, because the core grid holds only
whole cells and a window is any height it likes. `pardes_tagline_band_offset`,
`pardes_topbar_pane_border_px` and `pardes_topbar_pane_border_rgb` carry it over
the C ABI as PHYSICAL pixels: the host multiplies its points by the backing
scale going in and divides coming out, which is the snapping `Metrics` already
does for the cell, and is what keeps a one-pixel rule one pixel instead of a
two-pixel smear.

**The watch.** `file_watch.zig` was one mark/reconcile transaction over
`inotify`, so the tty shell, the SDL window and the detached daemon all watched
nothing off Linux: an edit made outside pardes never reached the pane, and a PDF
replaced on disk kept rendering the old inode. It is the same transaction over
two kernels now — `init`, `wait`, `stop`, `drain`, `markDir` and `unmarkDir` are
still the whole of it, and the hosts wait on a kqueue and poll it exactly as
they did the old descriptor. A macOS mark is TWO filters, because a kqueue
directory filter reports its entries changing and never a write to a file
already inside it: the parent mark follows rename-over saves, `markFile` catches
in-place writes, and `remarkFile` re-arms the file filter once a rename has moved
the inode. That is the same pair the AppKit host's DispatchSources already used
for the same reason. Directory marks are deduplicated here by device and inode,
because each `EVFILT_VNODE` filter needs a descriptor of its own and inotify did
that deduplication itself; `stop` and `drain` wake through the one `EVFILT_USER`
filter, since a kqueue cannot simply be read the way an inotify descriptor can.

**The effects.** The three `crt.ci.metal` entry points are
`extern "C" [[stitchable]]`. `CIKernel.kernels(withMetalString:)` compiles that
source at runtime, looks for stitchable functions, and rejects the WHOLE source
with "cannot find a valid stitchable Metal function in the source" when it finds
none — so `ScenePostprocessor.init?` returned nil and every scene effect and
panel transition silently degraded to the plain CoreText draw. The
`effect_sources.zig` test pins the exact spelling of all three, and
`draw-effect` in the e2e suite catches the degradation rather than the spelling.

Beside them, the offscreen harness owes the core a PRESENTATION. Its window is
borderless and never ordered front, so AppKit runs no display cycle and
`pardes_frame_presented` — whose only caller is `draw(_:)` — never fired. The
core holds pointer gestures inert while a layout mutation has not reached a
backend, which for an unpresenting harness is the rest of the script: the first
pane a script opened silently killed every later click, drag and Look. So
`readFrame` presents what it just rendered, into a bitmap nobody reads.
`PARDES_CHROME` also looks under `/Applications`, where a browser's executable
lives inside an application bundle and never on `PATH`. The macOS goldens are
regenerated; docs/macos.md, config.md, detached.md, web.md and the design PDF
follow.

2 files changed, 20 insertions, 15 deletions

diff --git a/src/detached/client.zig b/src/detached/client.zig
index 4769cf25..fc39164b 100644
--- a/src/detached/client.zig
+++ b/src/detached/client.zig
@@ -1129,9 +1129,20 @@ test "detached session: a frontend that stops reading is dropped, not waited for
     for (0..24) |_| {
         if (!h.session.clients[1].attached) break;
         host.vtable.push_set_clipboard.?(host.ctx, text);
-        try h.pump();
-        try good.wait(5);
-        while (try good.next()) |_| {}
+        // Read `good` back to EMPTY before the next mirror, rather than
+        // pumping once and taking whatever one write fitted. One pump moves at
+        // most one socket buffer, and that buffer is 8 KiB here
+        // (`net.local.stream.sendspace` on Darwin) against Linux's 208 KiB —
+        // so a single pump per 256 KiB mirror leaves the READING frontend
+        // falling behind by a quarter megabyte a round and closes it for a
+        // backlog it never caused. The claim under test is that a frontend
+        // that reads survives, so it has to actually finish reading.
+        for (0..256) |_| {
+            try h.pump();
+            try good.wait(5);
+            while (try good.next()) |_| {}
+            if (h.session.clients[0].out.items.len == 0) break;
+        }
     }
     // Dropped rather than queued without bound, and rather than the core
     // blocking on it.
diff --git a/src/detached/server.zig b/src/detached/server.zig
index 7d286177..2c4df87b 100644
--- a/src/detached/server.zig
+++ b/src/detached/server.zig
@@ -141,7 +141,6 @@
 const std = @import("std");
 const builtin = @import("builtin");
 const libc = std.c;
-const linux = std.os.linux;
 const posix = std.posix;
 const pardes = @import("../pardes.zig");
 const host_api = @import("../host.zig");
@@ -1221,17 +1220,13 @@ pub const Session = struct {
     /// then keeps its own record of what was asked and simply never gets a
     /// reload, which is what a host with no watcher has always done.
     ///
-    /// NONBLOCK because this descriptor is drained from `poll`, not from a
-    /// thread parked in `read` (tty.zig `watchFiles`): `drainInotify` must be
-    /// able to stop.
+    /// `polled` is true because this descriptor is drained from `poll`, not
+    /// from a thread parked in a wait (tty.zig `watchFiles`): `drainInotify`
+    /// must be able to stop. On linux that is IN_NONBLOCK; on macos a kqueue
+    /// needs nothing, since the timeout argument to `kevent(2)` decides.
     fn inotify(s: *Session) c_int {
         if (s.inotify_fd >= 0) return s.inotify_fd;
-        // The whole body is inside the comptime branch so that neither
-        // `inotify_init1` nor `linux.IN` is even analysed on a platform that has
-        // no inotify — the same shape file_watch.zig's `watchPath` uses.
-        if (comptime builtin.os.tag == .linux) {
-            s.inotify_fd = libc.inotify_init1(linux.IN.CLOEXEC | linux.IN.NONBLOCK);
-        }
+        s.inotify_fd = file_watch.init(true);
         return s.inotify_fd;
     }
 
@@ -1252,8 +1247,7 @@ pub const Session = struct {
     /// frontend got a frame per round it could not use. The fd is IN_NONBLOCK
     /// (`inotify`), so the loop ends on EAGAIN.
     fn drainInotify(s: *Session) void {
-        var buf: [4096]u8 = undefined;
-        while (libc.read(s.inotify_fd, &buf, buf.len) > 0) s.check_files = true;
+        if (file_watch.drain(s.inotify_fd)) s.check_files = true;
     }
 
     /// Reconcile every marked pane and the theme file. Called at the END of