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

[lesnik512]

Rewrites docs/adr/ from 3 records (961 words) to 3 (473 words), one paragraph each, in the shape ADR-FORMAT.md asks for. No record was dropped or merged, so the numbering and the slugs are unchanged and no citation had to move. 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 no connection ContextProvider	kept	0001	The boundary a sibling integration's reader will question; the provider would be public surface, so adding one later is a one-way door. Marginal on "hard to reverse" and the closest call of the three.
0002 no auto-inject sweep	kept	0002	Deliberate deviation from the obvious setup_di convenience, with a concrete rejected alternative that is still tempting.
0003 per-job child teardown is reference-counted	kept	0003	Reverting to the ownership boolean reintroduces a silent use-after-teardown; tests/test_jobs.py cites the record for exactly that.

Per record: 296 to 149 words, 272 to 155, 393 to 169. Removed from all three: the **Decision:** restatement of the title, the **Revisit trigger:** heading. 0001's and 0003's triggers are the real boundary of those decisions and survive as a clause in the paragraph; 0002's did not and was cut.

Citations

None moved. tests/test_jobs.py:238 still cites docs/adr/0003-per-job-child-teardown-is-reference-counted.md and docs/agents/domain.md still cites ADR-0001 and lists the first two filenames in its file-structure block; all resolve unchanged. The repo-wide sweep finds no reference to a number or a path that no longer exists.

Stale claims corrected

Checked against arq 0.28.0 in the lockfile and against modern_di_arq/main.py.

• 0002 described an arq.func(...) entry as carrying "its own name/timeout/retry configuration". arq.worker.Function is a six-field record: name, coroutine, timeout_s, keep_result_s, keep_result_forever, max_tries. Result retention was missing from the list, and it is part of what a sweep would have to rebuild. Now "name, timeout, result retention and retry settings" and "any of its six fields".
• 0002 said a sweep would have to handle an import string "without importing anything, since arq's own resolution of it is what decides which module is loaded and when". arq.func calls import_string eagerly and Worker.__init__ maps func over functions, so resolution happens at worker construction, not at job time. The conclusion holds; the wording now says what actually happens.
• 0001 claimed modern-di-celery and modern-di-typer "decline for the same reason". That is a claim about two other repositories and cannot be verified from this one, so it was dropped rather than carried forward.
• Verified and kept as written: 0001's ctx composition ({**self.ctx, **job_ctx} with job_id, job_try, enqueue_time, score, plus redis seeded by the worker) and the function.coroutine(ctx, *args, **kwargs) call site; 0003's depth key, its 1 to 0 close, the await-free transitions, and test_concurrent_inject_fanout_shares_one_child.

Gates

pytest -q tests/test_adr_citations.py (7 passed), the full suite (29 passed, nothing skipped: Redis was reachable, so the burst-worker tests ran), ruff format --check, ruff check --no-fix, eof-fixer . --check, and lychee --offline over the three records plus CONTEXT.md and AGENTS.md (0 errors). No Python was touched; the two ruff gates ran over the tree anyway.