reMarkable 2 support: windowed qtfb mode, arm32 fixes, one-command installer

[m4ndolore]

What this adds

Support for the reMarkable 2 as a second device, tested end-to-end on rM2 hardware (OS build 20260612, xovi + AppLoad v0.5.3). The rM2 runs the windowed (AppLoad/qtfb) backend only — takeover is inherently Paper Pro (libqsgepaper).

• rm2 cargo feature: 1404×1872 geometry, FBFMT_RM2FB, and the rM2's rotated wacom digitizer transform (Rot270 mount, inverted long axis — verified against libremarkable and on hardware).
• arm32 fix, affects any 32-bit device (incl. rM1): evdev input_event is 16 bytes on 32-bit userspace, not 24 — the parsers now use the kernel ABI for the target.
• arm32 fix, qtfb protocol: ServerMessage's payload union contains a size_t, so it sits at offset 8 with a u64 size on 64-bit servers but offset 4 with a u32 size on arm32. The client was reading the SHM key from the size field on rM2 (instant ENOENT). Parsing is now arch-correct for both.
• Cross-build: build-rm2.sh produces a static armv7-musl binary via cargo-zigbuild — no SDK needed for windowed mode.
• One-command installer (scripts/install-rm2.sh): rM2 needs no developer mode, so this goes from factory tablet to working diary — SSH key install, device check, xovi + AppLoad from official arm32 releases, xovi-tripletap persistence, riddle bundle, oracle config with provider-aware defaults, end-to-end --oracle-test. Fills the gap that remagic (Paper-Pro-only) leaves for rM2 owners.
• Docs (docs/rm2-setup.md): full walkthrough including the rM2 SSH quirks we hit in the field (dropbear 2020.81 hangs up if an RSA host key is negotiated — pin ed25519; stale known_hosts entries force exactly that failure).
• Feel tuning: RIDDLE_FLUSH_MS / RIDDLE_IDLE_MS env vars (windowed ink-update coalescing was the dominant tunable latency; the pause-to-commit is taste).
• Reply placement: Tom now writes below the user's ink when there's room — in windowed mode the drink dissolve leaves e-ink ghosting that can't be scrubbed without a flash, and the reply used to land on top of it.

Honest state on rM2

Works end to end: write → drink → oracle (OpenRouter, ~2.4 s to first ink) → streamed reply → fade. Two limits are structural to windowed mode and worth knowing:

• Ink latency is visibly above the stock notes app — every stroke rides shared memory → xochitl compositor → SWTCON. Great demo, not a place to write at length.
• Replies render fast-waveform gray, and client waveform-mode requests appear to be ignored by the compositor path on rM2. A per-update waveform hint in the qtfb protocol (an AppLoad change) would fix this for every qtfb client; happy to raise it with asivery if there's interest.

Where we're taking it (context, not part of this PR)

We looked at two roads to production-quality writing on rM2: (A) a takeover backend on rmkit's reverse-engineered SWTCON — feasible, unclaimed, may pursue later; (B) keeping stock xochitl notes as the writing surface and summoning the oracle onto the page. We're building B in our fork for knowledge-management use. This PR is the shared foundation for either.

Testing

• rM2 hardware: full flow, gestures, oracle streaming, window close/reopen.
• Paper Pro: aarch64-unknown-linux-gnu windowed build verified compiling (no takeover regression; no PP hardware here to run it).

Release suggestion: publish a riddle-appload-arm32.zip alongside the aarch64 bundle — build-rm2.sh assembles exactly that layout.

Field-tested and driven by @m4ndolore.

🤖 Generated with Claude Code

[m4ndolore]

UX from rM2 field testing — pushed to this branch (030e7cb), all from a couple of weeks of daily use:

• Send rule: a long flat stroke ruled under the words — like signing off a diary entry — sends the page deliberately. It has to span ~60% of the written text's width, gets absorbed off the page, and never reaches the oracle. Setting RIDDLE_IDLE_MS=0 disables the idle auto-send so the rule is the only trigger; in practice this felt much better than waiting for the timeout to fire mid-thought.
• RIDDLE_PERSONA_FILE (or a persona.txt next to the binary): swap Tom for any system prompt without rebuilding.
• RIDDLE_FONT_FILE: drop any TTF/OTF next to the binary to change the reply hand; unset keeps the embedded Dancing Script.
• RIDDLE_REPLY_WIDTH (default 4): thicker reply ink reads noticeably darker on the fast e-ink waveforms the windowed mode is stuck with.

All opt-in via env vars; defaults are unchanged except the reply pen width. Documented in oracle.env.example.