Initial commit: Naut-Torrent — from-scratch 10 GbE BitTorrent client

A maintainable, extensible BitTorrent client (C11, Linux/io_uring) targeting
10 GbE saturation. All torrent functionality is built from scratch; liburing
is the only linked third-party dependency on the data path.

Implements Phases 1-7 of the roadmap:
- core: page-aligned buffer pool, MPMC/Treiber queues, bitfields, worker pool
- crypto: SHA-1/256 (SHA-NI + scalar), Merkle (BEP-52), RC4 (MSE)
- bencode/metainfo: zero-copy parser, v1/v2/hybrid .torrent + magnet
- peer: sans-IO wire codec, MSE/PE handshake state machine, BEP-10, ut_metadata, PEX
- piece/storage: block-level multi-peer engine, rarest-first + endgame,
  per-file completion events + single-file relocate (move-as-you-finish)
- tracker/dht: HTTP + UDP (BEP-15) trackers, BEP-5 KRPC iterative lookup
- platform: io_uring reactor (SQPOLL, registered buffers, SEND_ZC)
- surface: versioned RPC, native plugin ABI, sandboxed Lua scripting, nautd/nautctl

Verified against libtorrent (single/multi/hybrid, MSE, magnet-via-DHT, swarm);
unit + interop tests green; ASan/UBSan/TSan clean. Scripting reference in
docs/scripting.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

include/naut/uring.h

@@ -0,0 +1,56 @@
+/* uring.h — io_uring lifecycle (the kernel-facing seam, part 2).
+ *
+ * Thin ownership wrapper over liburing so the rest of the engine never calls
+ * io_uring_* directly — that keeps the "Linux-only now, portable later" seam
+ * intact (a future epoll/kqueue backend implements the same reactor contract).
+ * Each reactor thread owns exactly one naut_ring for both network and disk.
+ */
+#ifndef NAUT_URING_H
+#define NAUT_URING_H
+
+#include "naut/common.h"
+#include "naut/buf.h"
+#include <liburing.h>
+
+typedef struct naut_ring {
+    struct io_uring ring;
+    bool sqpoll;
+    bool send_zc;
+    bool msg_ring;
+    bool buffers_registered;
+    bool recv_fixed;
+} naut_ring;
+
+/* entries: SQ depth (rounded up to a power of two by the kernel).
+ * sqpoll: dedicate a kernel thread to submission polling — removes the
+ * io_uring_enter syscall from the hot path once the queue is warm (needs
+ * CAP_SYS_NICE or /proc/sys tuning on some setups; falls back if it can't). */
+naut_err naut_ring_init(naut_ring *r, unsigned entries, bool sqpoll);
+naut_err naut_ring_init_cpu(naut_ring *r, unsigned entries, bool sqpoll,
+                            int sqpoll_cpu);
+void     naut_ring_close(naut_ring *r);
+
+/* Detect kernel support for the features the data path relies on. Logs a
+ * summary; returns NAUT_ERR_NOSYS if a hard requirement is missing. */
+naut_err naut_ring_probe(naut_ring *r);
+
+/* Register the pool's contiguous slab as fixed buffer index 0. Registration
+ * may fail under a low RLIMIT_MEMLOCK; callers can continue in degraded mode. */
+naut_err naut_ring_register_bufpool(naut_ring *r, const naut_bufpool *pool);
+void     naut_ring_unregister_buffers(naut_ring *r);
+
+/* Prepare a send using SEND_ZC when supported and requested, otherwise normal
+ * SEND. Returns true when the caller must retain the buffer until a CQE with
+ * IORING_CQE_F_NOTIF arrives. */
+bool naut_ring_prep_send(naut_ring *r, struct io_uring_sqe *sqe, int fd,
+                         const void *buf, size_t len, int flags,
+                         bool prefer_zero_copy);
+/* Returns true when this recv was armed against the registered fixed buffer.
+ * Some kernels accept READ_FIXED but reject IORING_RECVSEND_FIXED_BUF on plain
+ * recv with -EINVAL; the caller should remember this per-operation so it can
+ * distinguish "fixed-buffer unsupported, retry unfixed" from a real error
+ * (rather than tearing the connection down). */
+bool naut_ring_prep_recv(naut_ring *r, struct io_uring_sqe *sqe, int fd,
+                         void *buf, size_t len, int flags);
+
+#endif /* NAUT_URING_H */