coverage-nightly has never completed on the yoga pool (4/4 red since 2026-09-09): each night a different unit test asserts the box, so the 95 % ratchet has no baseline #3185

What happened

coverage-nightly.yml (#3123: low-priority coverage run on yoga toward the 95 % target) concluded failure on every scheduled run since it moved to the yoga pool — 34325094997 (09-09), 34451014241 (09-10), 34575134766 (09-11), 34680781819 (09-12) — and on the 0.67.0 cut when dispatched by hand (34702490406, yoga-build2, 2026-09-12T15:31Z). Every time make coverage (Makefile:586) exits 2 because ONE unit test fails under llvm-cov, and the test is different each time:

• main, 09-12 schedule: backends::gpu::device::backward::tests::test_silu_backward_at_zero and …_length_mismatch panic at crates/aprender-compute/src/backends/gpu/device/backward.rs:1103 / :1120 (no adapter on the box — the class fix(compute): nine GPU tests panicked without an adapter — the nightly coverage run was RED six days on an environment fact (PMAT-1106) #3116 fixed for nine other tests).
• the cut aa61219 (which contains fix(compute): nine GPU tests panicked without an adapter — the nightly coverage run was RED six days on an environment fact (PMAT-1106) #3116): those two pass; instead benchmark::tests::test_decompression_faster_than_compression panics at crates/aprender-zram-core/src/benchmark.rs:522: Decompression 2.6 MB/s much slower than compression 11.4 MB/s — a throughput ratio asserted inside a --lib unit test, measured under coverage instrumentation on a shared box.

So the coverage number for 0.67.0 does not exist, and §5 P3 of APR-RELEASE-001 ("target = measured baseline + 2 points per ratchet toward 95 %") cannot start.

Why

Two classes, both known here: a test that asserts the hardware it runs on (#3129, #3135, #3184 are the same shape on the GPU jobs), and a wall-clock/throughput assertion in a unit test (the rule in memory and in #3131, which already moved one such benchmark behind bench-gates).

Ask

1. benchmark.rs:522 — move the ratio assertion behind the bench-gates feature (as fix(tests): eight spec_checklist targets read repo paths from the crate dir — 77 vacuous rows now read the tree (#3130) #3131 did), or assert correctness (round-trip bytes) and record the throughput.
2. backward.rs:1103/1120 — skip with a named reason when no adapter is present, counted as skipped (the fix(compute): nine GPU tests panicked without an adapter — the nightly coverage run was RED six days on an environment fact (PMAT-1106) #3116 treatment).
3. make coverage — run the suite --no-fail-fast so a single box-assertion does not zero the night's report, and print the coverage number even when a test fails (report-only is the job's declared contract).
4. Then record the first baseline in docs/build-ledger/ per APR-RELEASE-001 §5 P3.

Refs #3123, #3116, #3131, #2307, PMAT-1098 (0.67.0 train).