test: gate branch coverage, time concurrency benchmarks on persistent workers, pay down test debt (#606)

[lesnik512]

Closes #606.

Summary

Baseline on ffe365c: 963 tests, 100% line coverage, 99.90% branch coverage of modern_di (two partial branches). After: 965 tests, 100% line and 100% branch coverage, and CI now enforces both.

1. Branch coverage gate. just test-ci measures modern_di only (run.source) with run.branch = true. The two partial branches are closed: collect_errors gets case _: typing.assert_never(event) (excluded from the report via exclude_also; ty fails if an Event member loses its case), and a new test cancels close_async after an earlier item finalized, then closes again, which reaches CacheItem.close_async with nothing owed. The test-branch recipe is gone because test-ci now does what it did.
2. G14/G15/G15b. The harness keeps N worker threads alive for the whole benchmark and releases them with two barriers per batch, so thread start-up and join are out of the timed call. G15 compiles its resolvers once and its per-round setup only closes and reopens the container, so each round times creation alone. G15b reuses one root and builds the children inside the job. New G15c row: an empty job on the same pool. Every scenario now asserts on what the timed batch resolved (G15: every thread got the same K instances; G15b: N x K distinct instances). README gets G15b and G15c rows and a rewritten harness paragraph. tests/test_bench_report.py only parses the comparative tier and still passes.
3. Free-threading tests. Every Barrier and join in test_free_threading.py and the two race tests in registries/test_providers_registry.py has a timeout, joins are followed by an is_alive assertion, and the threads are daemons so a hang fails the test instead of the run. A new thread_race marker covers that module plus the three race tests named in the issue; the 3.14t step runs just test -m thread_race --count=50 without the redundant -W error::RuntimeWarning.
4. Private API in tests. All 27 _providers_registry.add_providers/register(...) calls in tests now use container.add_providers(...); every registered type matched the provider's bound_type, so they are equivalent. The blanket tests/** SLF001 ignore is narrowed to the 13 files that still read internals with no public equivalent.
5. Coverage config. With only modern_di measured, all 54 # pragma: no cover comments and the 12 "exercise the body for coverage" asserts in tests are removed. run.concurrency, run.omit and addopts = "" are dropped.
6. Weak assertions. Removed the duplicate simplefilter("error") blocks; the partial-creator test now asserts the warning on 3.11-3.13 and its absence on 3.14+; the inspect.signature patch test asserts the patched call happened.
7. Lint. Dropped the ISC001 ignore.
8. CI. release.yml uses astral-sh/setup-uv@v8.2.0 like the other workflows. New allowed-to-fail matrix entries for 3.15 and 3.15t (continue-on-error), and the free-threaded steps now key on a trailing t. uv fetches cpython-3.15.0rc1; both builds pass the full gated suite locally after adding a repr(UNSET) test, which was only covered incidentally before 3.15.
9. AGENTS.md now says the lychee links job checks local links in every Markdown file.

Design decisions

• Measuring modern_di instead of . (the issue's exclude_also alternative) because branch coverage over tests reports 139 partial branches that are coverage.py artefacts of class X: ... bodies, and the pragmas existed only to satisfy line coverage of test code. The cost: a test helper that never runs is no longer reported.
• The 16 _cache_registry.cached_count() asserts are untouched; the registry refactor in Refactor and perf: registry hooks, container-owned cache and context, cheaper child build #605 replaces them.
• Free-threading stress step runtime on 3.14t, measured locally: 0.70 s of pytest time (1.2 s wall) before, 3.2 s (4.0 s wall) after. Most of the increase is collecting 48,000 repeated items to select 250.
• Local GIL numbers (3.14, median, 1/2/4 threads): G15 went from 232/280/382 us to 39/61/106 us, G15b from 225/319/499 us to 62/128/264 us; G15c is 11/24/53 us. The benchmark IDs are unchanged, so the first CI comparison will show G14/G15/G15b as large improvements.

Open questions

• The per-file SLF001 list will drift as Refactor and perf: registry hooks, container-owned cache and context, cheaper child build #605 removes private reads; ruff does not report an unused per-file ignore, so stale entries need a manual sweep.

Test plan

• just lint, just lint-ci
• just test-ci (965 passed, 100% line and branch)
• Gated suite on 3.11, 3.12, 3.13, 3.14t, 3.15rc1 and 3.15rc1t locally
• just test -m thread_race --count=50 on 3.14t and 3.15t
• Guard benchmarks on 3.14 and 3.14t; tests/test_bench_report.py
• mkdocs build --strict, just adr-check
• CI green

[github-actions]

Benchmark

Details

Benchmark suite	Current: cf0caa8	Previous: ffe365c	Ratio
benchmarks/test_guard_by_type.py::test_g16_resolve_by_type	3950185.9431616464 iter/sec (stddev: 7.745686381423656e-9)	4650912.939324458 iter/sec (stddev: 1.706440348160392e-8)	1.18
benchmarks/test_guard_by_type.py::test_g17_resolve_by_type_large_registry	3959333.683742874 iter/sec (stddev: 6.0933195179902515e-9)	4564403.042262547 iter/sec (stddev: 2.130887895421523e-8)	1.15
benchmarks/test_guard_cold.py::test_g8_cold_first_resolve	22511.995070418634 iter/sec (stddev: 0.00004111685903610907)	18672.690009299124 iter/sec (stddev: 0.000040957900608557336)	0.83
benchmarks/test_guard_cold.py::test_g8b_cold_first_resolve_cached	17528.177338987985 iter/sec (stddev: 0.00014840046722346283)	14435.165757881416 iter/sec (stddev: 0.00015649027293176276)	0.82
benchmarks/test_guard_concurrency.py::test_g14_concurrent_cached_hit[1]	512.1402568351706 iter/sec (stddev: 0.000057403048352041066)	606.7458247729055 iter/sec (stddev: 0.00006403397442405826)	1.18
benchmarks/test_guard_concurrency.py::test_g14_concurrent_cached_hit[2]	499.36872452706854 iter/sec (stddev: 0.00008212224430016972)	566.5625364040568 iter/sec (stddev: 0.00006131135700297531)	1.13
benchmarks/test_guard_concurrency.py::test_g14_concurrent_cached_hit[4]	477.667401533961 iter/sec (stddev: 0.00003973808132966421)	512.1861640209803 iter/sec (stddev: 0.00005466146361602645)	1.07
benchmarks/test_guard_concurrency.py::test_g15_concurrent_first_resolve[1]	12361.0947728639 iter/sec (stddev: 0.000010075445713161572)	1756.8118292460897 iter/sec (stddev: 0.0001580997113539773)	0.14
benchmarks/test_guard_concurrency.py::test_g15_concurrent_first_resolve[2]	7152.323327958439 iter/sec (stddev: 0.000018633341471109837)	1424.340746556056 iter/sec (stddev: 0.00017645930230460275)	0.20
benchmarks/test_guard_concurrency.py::test_g15_concurrent_first_resolve[4]	3712.139770477795 iter/sec (stddev: 0.00003179413836037132)	999.3575047320206 iter/sec (stddev: 0.00019075327136160947)	0.27
benchmarks/test_guard_concurrency.py::test_g15b_concurrent_first_resolve_sibling_children[1]	8715.81792393906 iter/sec (stddev: 0.000009341772940330055)	1776.3768085999434 iter/sec (stddev: 0.00019073682152782808)	0.20
benchmarks/test_guard_concurrency.py::test_g15b_concurrent_first_resolve_sibling_children[2]	3609.1900445262654 iter/sec (stddev: 0.000020340743363126515)	1244.4859807766236 iter/sec (stddev: 0.00018810472467390178)	0.34
benchmarks/test_guard_concurrency.py::test_g15b_concurrent_first_resolve_sibling_children[4]	1828.0617997614563 iter/sec (stddev: 0.000035023646802588695)	723.4024060255196 iter/sec (stddev: 0.0014598867005660315)	0.40
benchmarks/test_guard_concurrency.py::test_g15c_worker_pool_floor_control[1]	40243.39203514915 iter/sec (stddev: 0.000003460489734697166)
benchmarks/test_guard_concurrency.py::test_g15c_worker_pool_floor_control[2]	15869.07379347525 iter/sec (stddev: 0.000013541609614824049)
benchmarks/test_guard_concurrency.py::test_g15c_worker_pool_floor_control[4]	6608.72321723006 iter/sec (stddev: 0.00002684098432508923)
benchmarks/test_guard_lifecycle.py::test_g6_build_child_container	1204208.6490166876 iter/sec (stddev: 4.1114870543639516e-8)	1219360.5929127778 iter/sec (stddev: 4.021730185866024e-8)	1.01
benchmarks/test_guard_lifecycle.py::test_g6b_build_child_container_auto_scope	1088356.686696722 iter/sec (stddev: 2.1866267986318034e-8)	1165899.4125525632 iter/sec (stddev: 2.7625866083506307e-8)	1.07
benchmarks/test_guard_lifecycle.py::test_g7_request_lifecycle_batch	2533.6686100802854 iter/sec (stddev: 0.00001139311912002462)	2236.288706117797 iter/sec (stddev: 0.00001689205206591054)	0.88
benchmarks/test_guard_lifecycle.py::test_g7c_event_loop_floor_control	57115.703312321224 iter/sec (stddev: 0.0000016443527337652109)	67638.75445199345 iter/sec (stddev: 0.000001407229405623694)	1.18
benchmarks/test_guard_lifecycle.py::test_g7b_request_cycle_sync	310800.7165076886 iter/sec (stddev: 1.631130904295121e-7)	300382.96424878866 iter/sec (stddev: 9.599708707416182e-8)	0.97
benchmarks/test_guard_lifecycle.py::test_g13_teardown_at_scale	40906.247841908604 iter/sec (stddev: 0.0000017852769299060014)	41864.52786889268 iter/sec (stddev: 0.0000019744511004598563)	1.02
benchmarks/test_guard_lifecycle.py::test_g13b_teardown_at_scale_async_no_finalizers	459.329246138679 iter/sec (stddev: 0.001540877709414442)	469.4101679223813 iter/sec (stddev: 0.0010765366169991888)	1.02
benchmarks/test_guard_resolve.py::test_g1_transient_resolve	2370395.373655168 iter/sec (stddev: 2.719325034407428e-8)	2643693.8483676505 iter/sec (stddev: 3.622929811410649e-8)	1.12
benchmarks/test_guard_resolve.py::test_g2_cached_resolve	3826361.6959387376 iter/sec (stddev: 6.881111680281884e-9)	4533963.42125545 iter/sec (stddev: 1.7932439514885506e-8)	1.18
benchmarks/test_guard_resolve.py::test_g3_deep_chain	804851.499967102 iter/sec (stddev: 6.221465859700483e-8)	874097.5270822884 iter/sec (stddev: 1.3204596479966074e-7)	1.09
benchmarks/test_guard_resolve.py::test_g4_wide_resolve	474597.19690068497 iter/sec (stddev: 4.500930080375432e-7)	473981.3057505384 iter/sec (stddev: 6.531655111675316e-7)	1.00
benchmarks/test_guard_resolve.py::test_g5_cross_scope	2117031.0650465 iter/sec (stddev: 2.421736471148955e-8)	2097795.777795691 iter/sec (stddev: 1.4347586989334604e-7)	0.99
benchmarks/test_guard_resolve.py::test_g9_context_resolve	1599748.4555455411 iter/sec (stddev: 1.7486836072420347e-7)	1717618.5322124795 iter/sec (stddev: 1.6797887345620352e-7)	1.07
benchmarks/test_guard_resolve.py::test_g12_override_active_resolve	814359.9013097215 iter/sec (stddev: 3.83673928040184e-8)	912090.6924636723 iter/sec (stddev: 1.0201256745872507e-7)	1.12
benchmarks/test_guard_resolve.py::test_g18_alias_hop	3555794.3785150237 iter/sec (stddev: 9.776863197907824e-9)	4347042.296334007 iter/sec (stddev: 2.597541863228709e-8)	1.22
benchmarks/test_guard_validate.py::test_g10_validate_deep_chain	29421.917286472395 iter/sec (stddev: 0.00002187082673368506)	20347.922617500906 iter/sec (stddev: 0.00024790180939011935)	0.69
benchmarks/test_guard_validate.py::test_g11_validate_wide	17064.86456847256 iter/sec (stddev: 0.000027105569824449245)	13773.313381664582 iter/sec (stddev: 0.000027510882864299994)	0.81

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