diff --git a/docs/dependencies.md b/docs/dependencies.md
index 410e98e..d35ef91 100644
--- a/docs/dependencies.md
+++ b/docs/dependencies.md
@@ -54,7 +54,7 @@ else is on it.  A new workspace crate is added to it when it is created.
 | crates | may depend on |
 |---|---|
 | the protocol crates: npro-core, -h1, -ws, -h2, -h3, -quic, -wt, -mqtt, -http | nothing, except crates agreed below, each behind an opt-in feature, `no_std`, and with no build script |
-| npro-io, the IO adapter | std, and what its tls decision admits |
+| npro-io, the driver and the IO adapters | nothing without features; std, and the crates of each adapter's feature and of its tls decision ([io-model.md](io-model.md)) |
 | integrations with a runtime or framework, such as tokio | that runtime, in its own opt-in crate or feature, never a default |
 | npro-test, npro-fuzz, tools | what testing needs, admitted like anything else |
 | the libFuzzer targets in `fuzz/` | libFuzzer and what builds it, in a workspace of their own with its own `Cargo.lock` and `deny.toml`, so none of it enters the main workspace's graph |
@@ -154,9 +154,15 @@ in the commit adding the libFuzzer targets.
 
 ## Open
 
-- **tls**, for npro-io.  rustls is the obvious candidate, but it needs a
-  crypto backend, and its usual ones (`ring`, `aws-lc-rs`) carry C and
-  assembly with build scripts.  The alternatives are adapters to the
-  platform's tls or to OpenSSL.  This is decided when npro-io needs it
-  (phase 1g and after, in [port-plan.md](port-plan.md)), and recorded
-  here.
+- **tls**, for npro-io.  rustls goes inside the driver, behind a
+  record-layer trait.  rustls itself is `#![forbid(unsafe_code)]`; its
+  crypto provider is where the unsafe code is, and is the decision.  What
+  was found about `aws-lc-rs`, `ring`, `rustls-rustcrypto` and `graviola`
+  is in [io-model.md](io-model.md).  graviola leads for desktop and
+  server, but its licence (`Apache-2.0 OR ISC OR MIT-0`) is not MIT, and
+  it covers neither older CPUs nor microcontrollers.  Decided when the tls
+  commit comes, and recorded here.
+- **`mio`**, for npro-io's one-thread readiness loop (agreed in principle
+  2026-10-06), with `libc`, which has a build script, and `windows-sys`.
+  Admitted, with what `cargo deny` shows they bring, when that adapter is
+  written.
diff --git a/docs/io-model.md b/docs/io-model.md
new file mode 100644
index 0000000..5ff198a
--- /dev/null
+++ b/docs/io-model.md
@@ -0,0 +1,220 @@
+# The IO model
+
+How npro meets sockets, timers, tls and whatever scheduler the user has:
+tokio, threads with blocking calls, embassy on a microcontroller, or an
+event loop of their own.  Discussed and agreed with Andy on 2026-10-06;
+what is still open is listed at the end.
+
+## The rule
+
+**The protocol crates never call IO.  Anything that could block is a
+request the protocol returns, never an operation it calls.**
+
+Every protocol entry point is sync and non-blocking: it takes bytes and
+`now` as values, does its work, and returns.  That is the only kind of
+function that both sync and async code can call, so the same protocol
+code runs under any scheduler, and nothing in it is "coloured" sync or
+async.
+
+C's sansIO half calls out to IO: `want_write()`, `close()`, the
+deadline.  npro turns those calls around: `conn.poll_request()` returns
+`WantWrite`, `WantRead(bool)` or `Close(phase)`, and IO acts on the
+value.  An ops struct passed into the protocol would not give the same:
+its `read` must be blocking, which an async caller cannot use without
+parking a thread, or `async`, which a sync caller cannot use without a
+runtime.  A returned value also means no re-entry from a callback into
+the library, which was a source of trouble in C.
+
+So npro promises, and no feature changes it:
+
+- it never spawns a task or a thread on the user's behalf;
+- it never reads an ambient runtime or a global executor;
+- it puts no `Send` or `'static` bound on a connection.  A connection
+  is plain data, `Send` exactly when its storage and `Random` are.  The
+  one place `Send` appears is the cross-context handle (below), in the
+  adapters that offer one.
+
+A protocol crate may take a trait only for a service that is sync,
+cannot block, and does no IO: `Random`, and the tls record layer.  Time
+is a value, not a service.
+
+## Three layers
+
+1. **The protocol crates** (`npro-core`, `npro-h1`, `npro-ws`, ...):
+   `rx`, the `tx` pull, `deadline_passed`, `poll_request`, `poll_event`.
+   These exist.
+
+2. **The driver**, in `npro-io`, `no_std` without its `std` feature.  It
+   holds a connection, its buffers and its tls, and keeps the rules
+   every adapter would otherwise have to get right on its own: the order
+   of rx, requests, tx and deadline; a hangup at any point; a drain that
+   says whether output was delivered or dropped; tls stepped by one
+   function whose first step may already finish the handshake.  It does
+   no IO.  It tells the adapter what it wants, as a set, and the adapter
+   tells it what happened.  A first sketch, to be designed properly:
+
+   ```rust
+   enum Want<'a> {
+       Io { read: Option<&'a mut [u8]>, write: Option<&'a [u8]>, until: Option<Instant> },
+       Close,
+   }
+   // and back: read_done(now, n), write_done(now, n), timer(now), hangup(now)
+   ```
+
+   It must be a set, not one action at a time: a connection keeps
+   reading while it waits to write, or two peers each waiting to write
+   wedge.
+
+3. **Adapters**, which do the IO, each a short loop over the driver,
+   each behind its own `npro-io` feature.  Features are additive: each
+   adds an adapter and changes nothing else.
+
+   | adapter | feature | scale | needs |
+   |---|---|---|---|
+   | threads, blocking calls | `std` | a reader and a writer thread per connection; not for tens of thousands | std only |
+   | one-thread readiness loop | `mio` | many connections on one thread | `mio`, and through it `libc` (unix) and `windows-sys` |
+   | tokio | `tokio` | one task per connection, spawned by the user | tokio |
+   | embassy | `embassy` | connections fixed at build time (`pool_size`) | `embedded-io-async`, `embassy-time` |
+
+   **Your own loop** is supported too, as C supports foreign loops
+   (libuv, libevent, glib, sd-event): drive the driver from it directly.
+   The driver is the lowest public level, and needs nothing from npro-io's
+   adapters.
+
+Notes for each adapter, from the discussion:
+
+- **threads**: the driver behind a `Mutex` shared by the two threads,
+  which is allowed in an adapter but not in a protocol crate.  The writer
+  waits on a condvar for a write.  The read timeout comes from `until`.
+  `shutdown()` on the socket wakes a reader blocked in `read()`.  On
+  esp32 under esp-idf (std on FreeRTOS, sockets on lwIP) this runs
+  unchanged, but two thread stacks per connection cost RAM there; the
+  `mio` loop is the answer if that is too much.
+- **mio**: std has no `poll()`.  A readiness loop needs the OS's poll,
+  epoll or kqueue, and calling those is `unsafe`.  mio does that in about
+  170 marked places, and is what tokio stands on.  npro's own code stays
+  `#![forbid(unsafe_code)]`.
+- **tokio**: wait for readiness (`readable()`, `writable()`, a sleep
+  until `until`) and then do non-blocking `try_read` and `try_write`.
+  That way no buffer borrowed from the driver is held across an
+  `.await`, which is the same shape as a poll loop.  The future is
+  `Send` when the connection is; a `&mut dyn TxSource` held across an
+  `.await` must be `dyn TxSource + Send`, or it is not.  A compile-time
+  test asserts `Send`, so this cannot break silently.  Plugging into
+  hyper or axum is a separate question: npro is an alternative stack,
+  not a hyper backend.
+- **embassy**: buffers in `static` storage passed as `&'static mut`, not
+  inside the future, or every task becomes large.  To check: whether
+  embassy-net's `TcpSocket::read_with` and `write_with` hand a closure
+  the socket's own ring buffer.  If so, `rx` and the `tx` pull work on
+  it directly with no copy.
+
+### Writing from another task or thread
+
+`conn.send()` needs `&mut` access to the connection, so the
+application's code has to run where the connection is: on the adapter's
+thread, task or loop.  In C this is `lws_callback_on_writable()` plus
+`lws_cancel_service()`.  An adapter may offer a `Handle`: a bounded
+command queue plus a wakeup (a condvar, a waker, an eventfd).  The
+handle is where `Send + 'static` appears, and nowhere else.
+
+### What is written twice
+
+The user-facing conveniences, such as a blocking `client.get(url)` and
+an async `client.get(url).await`, are written once per adapter, thin,
+over the driver.  Macro crates that generate both versions from one
+source (`maybe_async` and others) are proc-macro dependencies, and are
+not used; the duplication is kept small and visible instead.
+
+## tls
+
+rustls is itself sans-IO.  It needs nothing from an event loop: its
+`TimeProvider` takes wall time, and randomness comes from its crypto
+provider.  It goes inside the driver, once:
+
+    socket -> rustls -> npro rx
+    npro tx pull -> rustls -> socket
+
+It is never wired in per adapter, so not `tokio-rustls`.  The unbuffered
+API (`UnbufferedClientConnection`, and its server equivalent) works on
+caller-owned buffers, as the rest of npro does.
+
+rustls is one implementation behind a small record-layer trait in the
+driver, not a fixed dependency, because the right tls differs by target.
+rustls needs `alloc`, and around 16 KiB of record buffer per direction.
+
+### The crypto provider
+
+rustls itself carries no unsafe code; the providers are where it lives,
+as in every tls stack, since constant-time and fast crypto ends up in
+assembly or intrinsics.  What was found on 2026-10-06:
+
+| provider | status | unsafe and build | platforms |
+|---|---|---|---|
+| `aws-lc-rs` | rustls' default; AWS | AWS-LC's C and assembly, compiled by its build script | wide |
+| `ring` | its author stepped back in 2025; RUSTSEC-2025-0007 said it was unmaintained, then was withdrawn once the rustls team took on its maintenance.  Commits are slow | mostly BoringSSL's C and assembly, compiled by its build script with a C compiler | wide |
+| `rustls-rustcrypto` | outside the rustls project; its README says not for production, incomplete, not audited | pure Rust over the RustCrypto crates, but not free of unsafe: their fast paths are `unsafe` intrinsics or `asm!` behind runtime cpu detection (`sha2` 0.10.9: 29 uses; `cpufeatures`: 9), which some crates can be built without, slower.  What remains even then is small and can't be avoided: `zeroize`'s volatile writes (15), so the compiler cannot drop a key wipe; `subtle`'s optimisation barrier (2); `getrandom`'s syscalls | anything Rust targets |
+| `rustls-graviola` / `graviola` | by Joseph Birr-Pixton (ctz), rustls' original author; its README says it is "very new".  0.4.1, 2026-06-24 | formally verified assembly from AWS's s2n-bignum for the ECC, RSA and ML-KEM arithmetic, intrinsics for AES and SHA; no C compiler.  Brings `cfg-if` and `getrandom` | **x86_64 and aarch64 only**, and only CPUs with the listed features (AES, SHA2, PMULL, NEON on aarch64, so no Raspberry Pi 4 or older; AVX2, ADX, BMI2, PCLMULQDQ and others on x86_64) |
+
+"Pure Rust" is not "no unsafe".  Memory safety is also not the main
+risk in crypto: timing side channels and correctness are, and that is
+what rustls-rustcrypto's README warns about.  The RustCrypto crates
+underneath are widely used on their own, and on a microcontroller they
+are likely to be in the tree anyway (`embedded-tls` uses them too).
+
+graviola is the best fit on desktop and server: no C, and verified
+assembly where the arithmetic is.  Its licence, `Apache-2.0 OR ISC OR
+MIT-0`, is not MIT, and `deny.toml` admits only MIT, so it needs a
+licence admitted as well as the crate.  It does not cover CPUs without
+those features, or any microcontroller.
+
+### esp32
+
+Two different worlds, and the chip matters too:
+
+- **esp-idf**: std on FreeRTOS.  The threads adapter works as it is.
+  tls can be rustls, since there is an allocator, but with a provider
+  that builds for the chip (graviola does not), or the IDF's own mbedTLS
+  with the chip's crypto hardware, through a separate binding crate that
+  nothing else depends on, as AGENTS.md allows for FFI.
+- **esp-hal and embassy**: `no_std`, usually with no allocator.  rustls
+  is out without `alloc`; `embedded-tls` is TLS 1.3 and client only,
+  over RustCrypto primitives.
+- **The chip**: the RISC-V parts (C3, C6) build with the normal Rust
+  toolchain; Xtensa (the original esp32, S2, S3) needs Espressif's own
+  fork of the compiler, which complicates pinning the MSRV in CI.
+
+Device tests run on Andy's hardware, and their output comes back by
+hand.
+
+## Proving the adapters behave the same
+
+The C transcripts that `npro-test` already replays against the protocol
+crates are replayed through the driver, and then through each adapter's
+loop over an in-memory transport, with identical results required.  An
+adapter that is not in that matrix is not claimed to behave the same.
+Then real sockets: npro's client against C's servers, C's clients
+against npro's servers, npro against itself, and autobahn both ways, over
+plain TCP first and over tls after.  These run locally until they pass,
+then become sai tasks.
+
+## Decided
+
+- 2026-10-06: the driver lives in `npro-io`, which is `no_std` without
+  its `std` feature; adapters are further additive features of it.
+- 2026-10-06: a `mio` readiness loop is supported, as well as threads,
+  tokio and embassy, and so is driving the driver from a loop of the
+  user's own.
+- 2026-10-06: tests against C lws run locally until they pass, then
+  move into sai.
+
+## Open
+
+- The tls provider on desktop: graviola is the leading candidate, which
+  needs its licence admitted.  What covers CPUs it does not, if anything.
+- tls on esp32, which follows from which world and which chip.
+- The driver's interface, designed by sketching all four adapter loops
+  against it, embassy's included, before it is fixed.
+- Admitting `mio`, `libc` and `windows-sys` (and whatever `cargo deny`
+  shows they bring) when the `mio` adapter is written.  `libc` has a
+  build script.
diff --git a/docs/port-plan.md b/docs/port-plan.md
index 1e5dba5..c73726f 100644
--- a/docs/port-plan.md
+++ b/docs/port-plan.md
@@ -364,15 +364,16 @@ page and the refusals the transcripts pin.
   - `npro-h1`: the h1 parser, chunked coding, and the h1 client and server
     transaction.
   - `npro-ws`: framing, handshake, close, and optional pmd.
-  - `npro-io`: std, sockets, the loop.  It comes later, and tls later
-    still.
+  - `npro-io`: the driver, `no_std`, and the adapters (threads, mio,
+    tokio, embassy), each an additive feature, `std` among them.  It
+    comes later, and tls later still ([io-model.md](io-model.md)).
   - `npro`: the facade.
   - `crates/npro-test` (`publish = false`): the transcript reader, the
     in-memory two-sided transport, the replay driver and the property-test
     helpers.
 - **Crate roots**:
   - every one carries `#![forbid(unsafe_code)]` and `#![no_std]`,
-    except `npro-io` and `npro-test`;
+    except `npro-test`, and `npro-io` when its `std` feature is on;
   - `alloc` only behind a feature;
   - `unexpected_cfgs` is denied in `[workspace.lints]`, along with
     `missing_docs` and the clippy set from AGENTS.md.
@@ -851,7 +852,16 @@ and other extensions are not ported.
 
 ### Phase 1g: minimal `npro-io`
 
-- std TCP, plain poll-style loop, no tls yet.  Enough to run:
+[io-model.md](io-model.md) is how npro meets sockets, tls and schedulers,
+agreed 2026-10-06.  For this phase:
+
+- the driver, `no_std`, replaying the transcripts over the in-memory
+  transport before any socket;
+- the threads adapter (`std`), std TCP, no tls yet.  The first form of
+  this plan said a poll-style loop, but std has no `poll()`; that is the
+  `mio` adapter, which waits for its dependency decision.  The minimal
+  examples (ws echo server and client, http server and client) come with
+  it.  Enough to run:
   - **autobahn**, both ways: `wstest -m fuzzingclient` against npro-io's
     ws echo server, and `-m fuzzingserver` against its echo client.  This
     is a conformance gate.  Where it runs is decided (2026-10-06): a sai
@@ -867,8 +877,11 @@ and other extensions are not ported.
     compare with;
   - **a differential run** of the port's client against C's
     `minimal-http-server` and ws echo, and the reverse.
-- tls: rustls is the obvious candidate.  It is a large dependency tree,
-  so it is a separate decision, not taken here.
+- the tests against C run locally until they pass, then become sai
+  tasks;
+- tls: rustls inside the driver, behind a record-layer trait, once plain
+  TCP passes against C; the same tests again over tls.  The crypto
+  provider is open: [io-model.md](io-model.md) has what was found.
 
 ### Later stages
 
@@ -896,5 +909,6 @@ not pending output; mux parked rx; the kept-warm joiner's status.
   lists the stage-1 set only.
 - Whether `Random` is `&mut dyn` or a generic parameter.  Generic avoids
   the vtable, but spreads a type parameter through every connection.
-- tls in `npro-io`.
+- tls in `npro-io`: the crypto provider, and tls on esp32
+  ([io-model.md](io-model.md), "Open").
 - Where autobahn runs in CI, and whether the C oracles are measured there (`sync-c-states.sh` and `sync-c-h1.sh` are run by hand).