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

[lesnik512]

Rewrites docs/adr/ from 2 records (303 words) to 2 (278 words), one paragraph each, in the shape ADR-FORMAT.md asks for. Both records pass the admission test, so nothing is dropped or merged and the numbering is unchanged; what changes is the shape. The **Decision:** lead and the **Revisit trigger:** trailer are gone from both, and each trigger that was the real boundary of its decision is folded into the prose instead of hanging off the end.

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 pure ASGI middleware	kept	0001	A deliberate deviation from the obvious path: a reader who sees raw scope/receive/send handling will ask why not BaseHTTPMiddleware, and the anyio-task and http-only answers are not visible in the code.
0002 child container stays internal	kept	0002	A public-surface boundary, and an explicit no: publishing an accessor is a name a major release has to keep working, and the reference-cycle reason for deleting the scope entry is nowhere in the code but the one-line comment.

No merges: 0001 decides the middleware's form, 0002 decides what the package exports. They meet at the ASGI scope dict but are separate trade-offs with separate reversal costs.

Citations

No citation moved, because no record was renumbered. The two that exist still resolve: tests/test_public_surface.py cites docs/adr/0002-child-container-stays-internal.md by full path in its INVARIANT: docstring, and docs/agents/domain.md names ADR-0001 plus both filenames in its file-structure block. The repo-wide sweep finds no reference to an old or dropped number.

Stale claims

None. Every concrete claim was checked against modern_di_starlette/main.py: _DIMiddleware is a plain ASGI callable, starlette_websocket_provider opens at Scope.SESSION, _CONTAINER_SCOPE_KEY is private and deleted in a finally, fetch_di_container(app) is the root-container accessor and has no per-connection counterpart, and @inject binds as a method (covered by tests/test_endpoints.py). The 0002 sentence about class-based endpoints kept its substance but dropped the issue reference, which the code and those tests now carry.

Gates

• pytest -q tests/test_adr_citations.py: 7 passed.
• pytest -q: 22 passed. The whole suite is unit-level and in-process via TestClient; no directory needed an external service, so none was skipped.
• ruff format --check (24 files already formatted) and ruff check --no-fix (all checks passed), run for completeness although no Python changed.
• eof-fixer . --check: clean.
• lychee --offline --no-progress over both ADRs, CONTEXT.md and AGENTS.md: 4 unique links, 2 OK, 0 errors, 4 excluded.