docs: input plan — Stage 5 blocker analysis (compositor + DRM)

Record the two concrete prerequisites for compositor input wiring, found by
reading the compositor source: (1) an event-loop restructure — it uses a
blocking accept loop, drains keys only reactively, and ClientState stores no
UnixStream, so no input source can push wl_keyboard to a client; (2) the
desktop cannot run without DRM/Mesa (no /scheme/drm/card0 on mini), so the
wiring is unvalidatable until the Mesa EGL/DRI port lands. Defines the exact
wiring for when those clear (read /scheme/evdev, set_vt_mode, real XKB keymap
fd, keycode = evdev - 8).

local/docs/INPUT-STACK-LINUX-ALIGNMENT-PLAN.md

@@ -390,6 +390,30 @@ duplication/ordering faults this structure removes by construction.
 - Boot check: a test client can `GrabDevice`, receive raw events, `Ungrab`, and the console resumes.
 
 **Stage 5 — Wire the compositor via seat (unblocks `redbear-full`).**
+
+> **Blocker analysis (2026-07-24, from reading `redbear-compositor/source/src/main.rs`):**
+> Two concrete prerequisites, neither runtime-testable today:
+> 1. **Compositor event-loop restructure.** The compositor uses a *blocking*
+>    `for stream in self.listener.incoming()` accept loop (`:1409`) and drains key
+>    events only *reactively* when a client poke arrives (`:2555`). Worse,
+>    `ClientState` (`:1145`) **does not store the client's `UnixStream`**, so no
+>    input source can push `wl_keyboard` events to the focused client. Real input
+>    requires: store per-client stream fds; multiplex the input fd + client
+>    sockets (poll loop or a dedicated input-reader thread); proactively send
+>    `wl_keyboard.enter/key/modifiers` to the focused client on input arrival.
+> 2. **The desktop cannot run without DRM/Mesa.** The compositor renders via
+>    `/scheme/drm/card0`; on mini/QEMU there is no DRM card
+>    ([[full-wayland-build-cascade-2026-07]]), so the compositor never starts and
+>    nothing here can be validated. This is desktop bring-up, gated on the Mesa
+>    EGL/DRI port — **not** a continuation of the input-core work.
+>
+> When those clear, the wiring is: compositor opens `/scheme/evdev` (evdevd's
+> Linux-evdev output — reuses evdevd's orbclient→evdev keycode translation, no
+> re-implementation), calls inputd `set_vt_mode(vt, graphics=true)` to suppress
+> the text console, ships a real XKB keymap fd (replacing the `/dev/null` stub at
+> `:3380`), and dispatches `wl_keyboard` to the focused surface (keycode = evdev
+> code − 8).
+
 - **Promote `seatd` to a local source package** (per §4b decision): flip `recipe.toml [source]` from
   `git = jackpot51/seatd` to `path = "source"`; remove the leftover nested `source/.git` (so it's a
   plain vendored tree like evdevd, not an embedded clone) and the stray `target/` in the recipe dir;