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

[lesnik512]

Rewrites docs/adr/ from 3 records (~740 words) to 2 (~300 words), one paragraph each, in the shape ADR-FORMAT.md asks for. Survivors are renumbered 0001-0002 in their original order and every citation moves with them.

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 dialog support is a submodule	kept	0001	Packaging and public-API boundary, genuinely costly to undo, and the rejected alternative (a separate modern-di-aiogram-dialog distribution, matching every other integration) is the obvious one.
0002 per-update context stays a literal	dropped	-	The context dict is internal to _DiMiddleware; routing it through integrations.bind is a three-line, zero-risk change, so the decision is not hard to reverse, and the alternative was rejected on length rather than on a constraint. No survivor relies on a rule it states, so nothing needed folding.
0003 no auto_inject for dialogs	kept, renumbered	0002	A user who sets auto_inject=True and finds dialog getters untouched will ask why; the answer is the structural constraint 0001 protects, so it is a real scope decision with a real alternative.

No merges: 0001 answers where dialog support lives, 0002 answers how far the sweep reaches. They share a constraint but not a question.

Citations moved

The only ADR citation in the repo is docs/agents/domain.md, which names ADR-0001 as its worked example of flagging a conflict; 0001 keeps its number, so the reference is unchanged. The file-structure block in the same file listed 0002-per-update-context-stays-a-literal.md and now lists 0002-no-auto-inject-for-dialogs.md. No Python, TOML or workflow file cites an ADR. The repo-wide sweep after editing finds no old or dropped number.

Stale claims corrected

• 0003 described the sweep as rewriting "the callback on every HandlerObject it finds". _inject_router skips the update observer and skips any callback integrations.is_injected already marks; 0002 now says so.
• 0003's "Dishka, the comparable integration, draws the line in the same place" was dropped rather than restated: it is a claim about another project that nothing in this repo verifies.
• Dropped 0002 claimed modern-di's kit "records aiogram as the documented Layer-1 outlier". The installed modern_di 3.4.0 integrations.py module docstring describes Layer 1 and Layer 2 but names no framework, and aiogram appears nowhere in that package. The record also dated the kit to modern-di 2.28 while this package now pins modern-di>=3,<4. Both go away with the record.
• 0001 said a separate distribution would have to re-export "this package's marker and private child-container key". dialog.py imports FromDI and the private helper _fetch_child_container, not the key constant; 0001 now says "its private container lookup".

Gates

Gate	Result
pytest -q tests/test_adr_citations.py	7 passed
pytest -q (full suite)	27 passed; no directory needs external services, so none was skipped
ruff format --check	25 files already formatted
ruff check --no-fix	all checks passed
eof-fixer . --check	clean
lychee --offline --no-progress over both ADRs, docs/agents/domain.md, CONTEXT.md, AGENTS.md	3 OK, 0 errors, 3 excluded

No Python changed, so the two Python gates only confirm the tree is still clean.