docs(adr): compress to 3 records in the domain-modeling format

docs/adr/0001-fake-broker-keeps-immediate-delivery.md

@@ -1,46 +1,12 @@
-# The fake broker keeps immediate delivery; no in-memory `TimerStore`
-
-**Decision:** `TestTimersBroker` keeps its current model — the fake producer encodes and dispatches
-every published Timer immediately to the handler via `process_message`, regardless of its Activation
-time. We do **not** add an `InMemoryTimerStore` adapter or a `TimerStore` `Protocol` for testing.
-
-## Context
-
-An architecture review proposed that "the fake fakes the wrong layer": put a real
-`InMemoryTimerStore` behind the `TimerStore` seam so the fake broker exercises real Claim / Lease /
-Commit and truthful inspection, instead of stubbing the Redis client and dispatching immediately.
-Three defects were claimed: (1) the fake duplicates the envelope encode; (2) `has_pending`,
-`get_pending_timers` and `cancel_all` are stubbed to canned empties ("lies"); (3) at-least-once and
-Lease semantics cannot be tested through the fake. The options weighed were: A — drive the *real*
-subscriber poll loop against an in-memory store; B — a real in-memory store for state with
-deterministic delivery; or drop it.
-
-## Decision & rationale
-
-Researching FastStream's own Redis broker (`faststream/redis/testing.py`) dismantled the premises:
-
-- **The encode is not duplicated.** FastStream's `FakeProducer.publish` also encodes via
-  `build_message` / `message_format.encode` — encoding is inherent to the fake-producer pattern, not
-  a smell. `FakeTimersProducer` mirrors it.
-- **The inspection stubs are not lies.** FastStream fakes *every* subscriber — including the polling
-  list subscriber, our closest analog — by bypassing the poll loop and calling
-  `handler.process_message` directly from the fake producer; it never runs the real loop in tests.
-  Under an immediate-delivery contract a published Timer has already fired and been removed, so
-  `has_pending → False`, `get_pending_timers → []` and `cancel_all → 0` are *truthful*, not stubbed
-  lies.
-- **Option A is non-idiomatic and fragile.** Running the real `_consume` loop under FastStream's
-  `TestBroker` means skipping `_fake_start` — forfeiting the handler-mock wiring every FastStream
-  test broker relies on — and driving an infinite 50 ms poll loop inside a unit test; a spike doing
-  so hung at teardown.
-
-That leaves only one way to make inspection report *pending* future Timers: stop delivering future
-Timers immediately. That is a **breaking change** to a public testing API — existing user tests
-publish a future Timer and expect their handler to fire without waiting — so it is rejected. With
-immediate delivery retained the inspection methods are already correct, and the proposal has no
-remaining defect to fix. `scheduled_timers` already lets users assert what was Scheduled.
-
-**Revisit trigger:** a concrete need to test Schedule / Pending / cancel semantics — a future Timer
-observed as Pending before it fires — through the fake. If that arises, prefer an **opt-in**
-`TestTimersBroker(..., respect_activation=True)` (default off, so non-breaking) that withholds
-not-yet-Due Timers; not a change to the default immediate-delivery contract, and not a real
-subscriber loop in tests.
+# The fake broker delivers every Timer immediately
+
+`FakeTimersProducer` encodes each published Timer and hands it straight to the matching subscriber's
+`process_message`, ignoring the Activation time, and `_patch_broker` stubs the Redis client so
+`has_pending` is False, `get_pending_timers` empty and `cancel_all` zero. Putting a real in-memory
+store behind the `TimerStore` seam, so the fake exercised Claim, Lease and Commit, was rejected:
+FastStream fakes every subscriber, its polling list subscriber included, by bypassing the poll loop,
+and under immediate delivery a published Timer has already fired, so those inspection results are
+truthful rather than canned. Driving the real `_consume` loop under `TestBroker` means skipping
+`_fake_start` and running a 50 ms poll loop inside a unit test; a spike doing so hung at teardown.
+Reporting a future Timer as Pending requires withholding it, which breaks every user test that
+publishes one and expects its handler to fire, so it could only arrive as an opt-in flag.

docs/adr/0002-no-per-call-middlewares-kwarg.md

@@ -0,0 +1,14 @@
+# No per-call `middlewares=` on subscribers, publishers and routes
+
+`TimersRegistrator.subscriber()`, `.publisher()`, `TimersRoute` and `TimersRoutePublisher` take no
+`middlewares=` keyword; `TimersBroker` and `TimersRouter` keep theirs, and broker or router scope is
+the only way to wrap timer handling. FastStream 0.7 removed `add_call(middlewares_=...)`, so the
+forwarding target no longer exists. Routing the kwarg to broker scope internally would have been a
+silent change of meaning, since broker middleware runs for every Topic, and re-implementing
+per-subscriber middleware locally would reproduce what upstream had just deleted while binding us to
+an internal we had stopped depending on. The package was at version `"0"` with no stability promise,
+so a hard break users see at import time was in policy. Behaving exactly like the brokers FastStream
+ships is this integration's value, which is also why the single-caller `create_publisher` and
+`create_subscriber` factories stay rather than being inlined into the registrator: they mirror
+FastStream's own Redis broker layout. If upstream restores call-scoped middleware the kwarg returns
+as a thin forward.

docs/adr/0003-typing-extensions-over-version-gated-imports.md

@@ -0,0 +1,13 @@
+# Backport 3.12-only typing via `typing_extensions`, not version-gated imports
+
+Lowering the supported floor from 3.13 to 3.11 broke exactly two constructs: the PEP 695
+`type RedisClient = ...` alias, a `SyntaxError` on 3.11, and `override` imported from `typing`,
+which landed in 3.12. The alias became a plain `typing.TypeAlias` annotation, which needs no
+backport; `override` is imported from `typing_extensions` unconditionally, never gated on
+`sys.version_info`, and `typing-extensions` is a declared direct runtime dependency rather than a
+pin borrowed from FastStream, which is free to drop it. A `sys.version_info` gate was rejected
+because 3.11 needs the package installed either way, so it buys only more code at every affected
+site and a second path for `ty` to check. Dropping `@override` was rejected because it is what
+catches a signature drift against FastStream's base classes, precisely the failure this integration
+is exposed to. When the floor rises to 3.12 both constructs gain stdlib spellings and the direct
+dependency goes with them.

docs/agents/domain.md

@@ -18,7 +18,7 @@ Single-context repo:
 ├── CONTEXT.md
 ├── docs/adr/
 │   ├── 0001-fake-broker-keeps-immediate-delivery.md
-│   └── 0002-keep-construction-factories.md
+│   └── 0002-no-per-call-middlewares-kwarg.md
 └── faststream_redis_timers/
 ```
 
@@ -32,4 +32,5 @@ If the concept you need isn't in the glossary yet, that's a signal: either you'r
 
 If your output contradicts an existing ADR, surface it explicitly rather than silently overriding:
 
-> _Contradicts ADR-0001 (fake broker keeps immediate delivery), but worth reopening because…_
+> _Contradicts ADR-0001 (the fake broker delivers every Timer immediately), but worth reopening
+> because…_