docs(adr): compress to 2 records in the domain-modeling format

[lesnik512]

Rewrites docs/adr/ from 2 records (535 words) to 2 (306 words), one paragraph each, in the shape the domain-modeling ADR format asks for. Both records pass the admission test, so nothing is dropped and nothing is merged; numbering and filenames are unchanged. What goes is the format: the **Decision:** lead, the rejected-alternative bullet lists, and the **Revisit trigger:** block. Same treatment as modern-python/compose2pod#102.

Evaluation

Each record was tested against the admission test: hard to reverse, surprising without context, the result of a real trade-off.

Old	Outcome	New	Reason
0001 metrics recorders not unified	kept	0001	Hard to reverse (a unification is a rewrite of both adapters), surprising (two switches on the same event names read as duplication a reader will try to "fix"), and a real trade-off (a declarative (kind, name, label set) table was weighed and rejected).
0002 free-threading is compat-only	kept	0002	Hard to reverse (a published classifier and a support claim are not quietly withdrawn), surprising (free-threaded support that deliberately adds no parallelism, plus a required env var), and a real trade-off (four alternatives were weighed, including exploiting the parallelism).

0002's revisit trigger is folded in rather than kept as a block: DISABLE_SQLALCHEMY_CEXT_RUNTIME=1 is the real boundary of the guarantee, not a scheduling note, so it is stated as the scope of the decision with #160 as the issue link. 0001's parity-contract-test bullet survives as a clause, since the rejection is the non-obvious half of the decision.

Citations

None move. The only ADR reference in the repo is the ADR-0001 example in docs/agents/domain.md, which still names record 0001 and the same subject; the file-structure listing in the same file is unchanged. The repo-wide sweep for adr/NNNN and ADR-NNNN across *.py, *.md, *.toml and *.yml finds no old or dropped number.

Stale claims

None. Every concrete claim was re-checked against the tree and holds: the two __call__ switches in faststream_outbox/metrics/{prometheus,opentelemetry}.py with no shared table; dispatched as three Prometheus operations (_received_total.inc, _received_size.observe, _in_process.inc) and an explicit OpenTelemetry no-op; no parity contract test in tests/; the freethreaded CI job in .github/workflows/_checks.yml with its DISABLE_SQLALCHEMY_CEXT_RUNTIME=1 env and sys._is_gil_enabled() assertion, pinned to 3.14t; the Programming Language :: Python :: Free Threading :: 2 - Beta classifier in pyproject.toml; the docs note in docs/introduction/installation.md; and #160 still open. One wording change, not a correction: the classifier is now named explicitly instead of called "a classifier".

Gates

Gate	Result
pytest -q tests/test_adr_citations.py	7 passed
pytest -q (full suite)	516 passed, 107 skipped - every skip is tests/test_integration.py / tests/test_client_contract.py reporting Postgres unreachable on the local DSN
ruff format --check	96 files already formatted
ruff check --no-fix	all checks passed
eof-fixer . --check	clean (exit 0)
lychee --offline --no-progress over both ADRs, CONTEXT.md, AGENTS.md	3 links, 2 OK, 0 errors, 1 excluded (the offline run skips the remote #160 link)

No Python was touched, so the ruff gates are informational here.

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	982	1.000	6.79	917	242	5000	5000	10000
consumer/w1/b100	986	1.000	6.83	919	243	5000	5000	10000
consumer/w2/b10	1101	1.000	6.81	966	242	5000	5000	10000
consumer/w2/b100	1341	1.000	6.73	955	243	5000	5000	10000
consumer/w4/b10	1040	1.000	6.76	969	262	5000	5000	10000
consumer/w4/b100	1578	1.000	6.86	1036	244	5000	5000	10000
consumer/w1/b100/tfbs100	5357	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2079	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls, exact; select_calls within +2; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).