refactor: tidy registry, wiring, factory and container internals (#575)

[lesnik512]

Closes #575.

Summary

Internal cleanup. Behaviour and public API stay the same, and the generated resolver templates in resolver_compiler.py are untouched.

• Providers registry. The three generation-checked stores now go through one _publish helper, generic over the memo's key and value types. drop_resolvers and _invalidate share _drop_resolvers.
• Wiring data clump. plan_for(provider) and WiringPlan.build(owner, registry=...) take the factory and read its parameters and kwargs themselves. The factory no longer hands its own private fields back alongside itself.
• WiringPlan. Dropped pure_provider (it was not static_kwargs) and the edges alias of provider_kwargs. Factory._can_call_positionally checks static_kwargs directly, and the graph walk reads provider_kwargs. AGENTS.md and two test docstrings now name provider_kwargs.
• CacheItem.settings is required. Only cached factories reach fetch_cache_item, so the settings and ... guards were dead outside tests. The duplicated finalize condition is now _pending_finalizer(), and fetch_cache_item takes Factory[typing.Any]. The cache registry tests now build items with CacheSettings(), and the "bare" item with no settings is gone from the close test.
• Factory.__init__. Parsing and checks moved into _parse_creator, _reject_unknown_kwargs and _reject_unresolvable_generics. ParsedCreator gained accepts_any_kwargs, so the unknown-kwarg check no longer calls inspect.signature a second time. The warnings use stacklevel=3 and still point at the Factory(...) line (checked by hand).
• Alias. _find_source is gone. _redirect_target is a plain lookup that returns None for a missing source, so no exception is used for control flow. _get_dependencies raises AliasSourceNotRegisteredError on None.
• Sentinels. is UNSET everywhere ty does not need to narrow (OverrideHandle.__exit__, both skip_creator_parsing/union-warning checks). isinstance(x, UnsetType) stays only where ty narrows the type afterwards, since is UNSET does not narrow.
• dependency_graph. The stateless DependencyGraph class is now the module functions walk and find_cycle_from. container.py calls them through the module, so tests patch dependency_graph.find_cycle_from and dependency_graph.walk.
• Group. typing.Any became object in __new__ and __init_subclass__, which removes both noqa: ANN401.
• Naming. SignatureItem.args is now member_types, matching ArgumentResolutionError.member_types. Factory._parsed_kwargs is now _params, matching ParsedCreator.params.
• Comments. Removed the two SLF001 rationale comments, the "inlined rather than a helper" comment and the coverage-tracer paragraph in _handle_recursion_error's docstring. The comments that stay are one line each.

Design decisions

• WiringPlan.build reads owner._params and owner._kwargs with an inline noqa: SLF001. A per-file ignore for wiring.py was the other option, but two reads did not seem worth one. Two precedence tests used to pass made-up parameter maps straight to build. They now set owner._params on a throwaway factory first.
• ParsedCreator.accepts_any_kwargs is also True when inspect.signature fails. The old check returned early in that case too, so no kwargs= key gets rejected for an uninspectable creator.

Bullets already done or not applicable

• ParsedCreator.return_type as a whole SignatureItem. Kept. Factory reads .member_types from it for the union-return warning, not only .arg_type.
• Exception __slots__. The repeated ("_base_message", "dependency_path") pair now appears only on ResolutionError. Moving it onto DependencyPathMixin fails with "multiple bases have instance lay-out conflict" next to RuntimeError. Dropping __slots__ makes instances weak-referenceable, which is a small behaviour change, so I left them as they are.
• _compile_alias duplicating the source lookup. It is in resolver_compiler.py, works on the registry rather than a container, and that file belongs to Refactor: resolver compiler templates and resolution-step building #574. Left alone.
• Line numbers, the registries becoming private container attributes and the container close rewrite were already handled by feat!: FinalizerError is an ExceptionGroup (#564) #583 to feat!: tighten Container's public surface (#565) #592.

Resolved questions

• CacheRegistry.fetch_cache_item keeps its typing.cast from CacheSettings | None, because only the cached template calls it. Passing the settings in from the compiler would mean editing resolver_compiler.py, which belongs to Refactor: resolver compiler templates and resolution-step building #574.

Test plan

• just lint and just lint-ci clean
• just test-ci: 617 passed, 100% line coverage
• New test_factory_warning_points_at_the_factory_call checks that both Factory warnings report the caller's file and line. It fails with stacklevel=2.
• Guard-tier benchmarks (pytest benchmarks/ --ignore=benchmarks/comparative --benchmark-only): 29 passed
• mkdocs build --strict clean
• CI green

[github-actions]

Benchmark

Details

Benchmark suite	Current: 139bc34	Previous: d989384	Ratio
benchmarks/test_guard_by_type.py::test_g16_resolve_by_type	2800759.9582160907 iter/sec (stddev: 1.306901301847443e-8)	3011454.7003564946 iter/sec (stddev: 1.1323228114990187e-8)	1.08
benchmarks/test_guard_by_type.py::test_g17_resolve_by_type_large_registry	2755316.058588567 iter/sec (stddev: 1.3273165229357577e-8)	3019957.023593222 iter/sec (stddev: 1.093424802027123e-8)	1.10
benchmarks/test_guard_cold.py::test_g8_cold_first_resolve	20530.035817689233 iter/sec (stddev: 0.00005018214516050216)	20349.205785269238 iter/sec (stddev: 0.0000523667240716527)	0.99
benchmarks/test_guard_cold.py::test_g8b_cold_first_resolve_cached	16291.624553202475 iter/sec (stddev: 0.0001188397023851702)	16003.317127720646 iter/sec (stddev: 0.00012320275507795883)	0.98
benchmarks/test_guard_concurrency.py::test_g14_concurrent_cached_hit[1]	363.8081338379358 iter/sec (stddev: 0.00006149524436734405)	394.762689858709 iter/sec (stddev: 0.00007995055136708627)	1.09
benchmarks/test_guard_concurrency.py::test_g14_concurrent_cached_hit[2]	345.31715073637986 iter/sec (stddev: 0.000058203081407507076)	351.6911874614162 iter/sec (stddev: 0.000454984947973727)	1.02
benchmarks/test_guard_concurrency.py::test_g14_concurrent_cached_hit[4]	322.6305153999816 iter/sec (stddev: 0.00005313844459959927)	345.1788240812406 iter/sec (stddev: 0.00006543668197917592)	1.07
benchmarks/test_guard_concurrency.py::test_g15_concurrent_first_resolve[1]	1928.3533180369154 iter/sec (stddev: 0.00016705275736381535)	1934.6799490405135 iter/sec (stddev: 0.000182087980571908)	1.00
benchmarks/test_guard_concurrency.py::test_g15_concurrent_first_resolve[2]	1490.3280688987932 iter/sec (stddev: 0.0001717478495164807)	1492.4575857536497 iter/sec (stddev: 0.0001797118465704635)	1.00
benchmarks/test_guard_concurrency.py::test_g15_concurrent_first_resolve[4]	1044.8563628122886 iter/sec (stddev: 0.00017920148930629566)	1050.1470415889385 iter/sec (stddev: 0.0001817990035889884)	1.01
benchmarks/test_guard_concurrency.py::test_g15b_concurrent_first_resolve_sibling_children[1]	1873.0512072831907 iter/sec (stddev: 0.00019811424111559198)	1910.6620291062904 iter/sec (stddev: 0.00017286707722268446)	1.02
benchmarks/test_guard_concurrency.py::test_g15b_concurrent_first_resolve_sibling_children[2]	1326.9494385600083 iter/sec (stddev: 0.0001763031471026032)	1141.5328994635493 iter/sec (stddev: 0.0013432424266359635)	0.86
benchmarks/test_guard_concurrency.py::test_g15b_concurrent_first_resolve_sibling_children[4]	766.3139621507365 iter/sec (stddev: 0.0011320296416828276)	847.901559872413 iter/sec (stddev: 0.00020861333395206277)	1.11
benchmarks/test_guard_lifecycle.py::test_g6_build_child_container	945413.2566797478 iter/sec (stddev: 4.411633039500867e-8)	973681.6763908018 iter/sec (stddev: 4.56938148133369e-8)	1.03
benchmarks/test_guard_lifecycle.py::test_g6b_build_child_container_auto_scope	880626.3631138019 iter/sec (stddev: 2.7197862219509484e-8)	885870.3741764443 iter/sec (stddev: 8.512144883629969e-8)	1.01
benchmarks/test_guard_lifecycle.py::test_g7_request_lifecycle_batch	2365.0806679514244 iter/sec (stddev: 0.000009746185481414945)	2418.3615689175267 iter/sec (stddev: 0.00003596881927590391)	1.02
benchmarks/test_guard_lifecycle.py::test_g7c_event_loop_floor_control	57681.47281610466 iter/sec (stddev: 0.0000025851155368536627)	53782.15896686228 iter/sec (stddev: 0.000006961250967947257)	0.93
benchmarks/test_guard_lifecycle.py::test_g13_teardown_at_scale	40337.80819895327 iter/sec (stddev: 0.000002111820985926339)	41835.43269737027 iter/sec (stddev: 0.0000033666901669700696)	1.04
benchmarks/test_guard_lifecycle.py::test_g13b_teardown_at_scale_async_no_finalizers	510.44935077391244 iter/sec (stddev: 0.000022310343707135817)	535.8588913626991 iter/sec (stddev: 0.00003599390005431521)	1.05
benchmarks/test_guard_resolve.py::test_g1_transient_resolve	2084065.1441326537 iter/sec (stddev: 2.542606520528666e-8)	1970315.5219002054 iter/sec (stddev: 3.602626171593177e-8)	0.95
benchmarks/test_guard_resolve.py::test_g2_cached_resolve	3107700.016694698 iter/sec (stddev: 1.4661207313646741e-8)	2988615.0205070125 iter/sec (stddev: 1.4195983575456663e-8)	0.96
benchmarks/test_guard_resolve.py::test_g3_deep_chain	694530.7423016601 iter/sec (stddev: 4.880839202154771e-8)	703914.1055850635 iter/sec (stddev: 4.0172620649808416e-8)	1.01
benchmarks/test_guard_resolve.py::test_g4_wide_resolve	439587.9355442461 iter/sec (stddev: 3.7525631912197367e-7)	456542.412562416 iter/sec (stddev: 3.6789550844346597e-7)	1.04
benchmarks/test_guard_resolve.py::test_g5_cross_scope	1889775.0884144164 iter/sec (stddev: 2.0078037678759696e-8)	1849364.1146450916 iter/sec (stddev: 2.60036094632861e-8)	0.98
benchmarks/test_guard_resolve.py::test_g9_context_resolve	1332733.425589795 iter/sec (stddev: 1.2704744037183457e-7)	1329519.1202467596 iter/sec (stddev: 1.532936950112306e-7)	1.00
benchmarks/test_guard_resolve.py::test_g12_override_active_resolve	661291.0755985831 iter/sec (stddev: 4.6476599453976785e-8)	699879.2743247835 iter/sec (stddev: 5.822605001118232e-8)	1.06
benchmarks/test_guard_resolve.py::test_g18_alias_hop	2884471.743470469 iter/sec (stddev: 9.953240555059758e-9)	3107097.9695588322 iter/sec (stddev: 1.0543047063911014e-8)	1.08
benchmarks/test_guard_validate.py::test_g10_validate_deep_chain	27188.67915592998 iter/sec (stddev: 0.000027909728576002846)	26119.65196646598 iter/sec (stddev: 0.000025633985370715704)	0.96
benchmarks/test_guard_validate.py::test_g11_validate_wide	15989.422421410625 iter/sec (stddev: 0.000025913601680724643)	15718.498294278044 iter/sec (stddev: 0.000026702657245432177)	0.98

This comment was automatically generated by workflow using github-action-benchmark.