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

[lesnik512]

Rewrites docs/adr/ from 4 records (~1,460 words) to 3 (~470 words), one paragraph each, in the
shape the domain-modeling ADR format asks for. The **Decision:** line, the ## Context and
## Decision & rationale headings and the **Revisit trigger:** section are gone; a revisit
trigger survives only where it is the real boundary of the decision. Survivors keep their original
relative order and are renumbered 0001-0003.

Evaluation

Each record was tested against the admission test: hard to reverse, surprising without context, the
result of a real trade-off. A record that fails on its own but states a rule a survivor relies on
was folded into that survivor as a sentence.

Old	Outcome	New	Reason
0001 fake broker keeps immediate delivery	kept	0001	Breaking change to a public testing API, and a reader who sees _patch_broker mocking the Redis client will propose the in-memory store that was already rejected.
0002 keep construction factories	dropped	-	Fails the test on two counts: inlining two pure constructor-bundling functions is the cheapest possible reversal, and the record decides not to do a cleanup rather than choosing between alternatives. Its rule, that the package mirrors FastStream's own Redis broker layout, is folded into the new 0002 as one sentence.
0003 no per-call middlewares= kwarg	kept, renumbered	0002	A removed public kwarg with a deliberate hard break, and both rejected alternatives (silent broker-scope routing, a local re-implementation) are non-obvious.
0004 typing-extensions over version-gated imports	kept, renumbered	0003	A declared runtime dependency plus an unconditional typing_extensions import that looks redundant on 3.13; the version gate and dropping @override were both real options.

No merges: the three survivors are about different boundaries.

Citations

Only two existed. docs/agents/domain.md names 0002-keep-construction-factories.md in its
file-structure example, now 0002-no-per-call-middlewares-kwarg.md, and cites ADR-0001 in the
conflict-flagging example, whose number is unchanged but whose parenthetical title was updated to
match the rewritten record. The intra-ADR [ADR-0001](0001-fake-broker-keeps-immediate-delivery.md)
link lived in the dropped record and went with it. No Python cites an ADR. The repo-wide sweep finds
no reference to an old or dropped number.

Stale claims

None. Every concrete claim was re-checked against the code and all of them still hold:
FakeTimersProducer.publish encodes through TimerMessageFormat.encode and calls
process_message directly, TestTimersBroker._patch_broker stubs zrangebyscore/zscore/the
cancel_all pipeline, scheduled_timers exists, TimerStore is a concrete class with no
Protocol and no in-memory sibling, the subscriber's default polling_interval is still 0.05,
neither TimersRegistrator.subscriber()/.publisher() nor TimersRoute/TimersRoutePublisher
takes middlewares= while TimersBroker/TimersRouter still do, RedisClient is a
typing.TypeAlias, override is imported unconditionally from typing_extensions in all three
modules that use it, typing-extensions>=4.12.0 is a direct dependency and requires-python is
>=3.11. Two details worth naming: the dropped record's premise was accurate (create_publisher
and create_subscriber really are single-caller pass-throughs), and datetime.UTC, the constant
0004 credited with fixing 3.11 as the floor, is still imported in response.py.

Gates

Gate	Result
pytest -q tests/test_adr_citations.py	7 passed
pytest -q (no-Redis suites: test_unit, test_fake, test_store_seam, test_tuning_defaults, test_adr_citations)	108 passed
ruff format --check	56 files already formatted
ruff check --no-fix	All checks passed
eof-fixer . --check	clean
lychee --offline over the three ADRs, docs/agents/domain.md, CONTEXT.md, AGENTS.md	3 links, 2 OK, 0 errors, 1 excluded

The integration suites (test_at_least_once, test_cancel, test_concurrency, test_delivery,
test_envelope, test_inspection, test_isolation, test_publish_api) need a live Redis and were
not run locally; CI covers them. No Python was touched, so the ruff and eof-fixer runs are
repo-wide rather than scoped to the change.