folded_health fixtures fail when the CI host itself is io-stalled — the real degradation probe leaks into tests that assume an undegraded node

[emooreatx]

On PR #566 (workflow/installer files only), clippy + test (ubuntu-latest) run 34229542137 failed four tests/folded_health.rs cases:

a_brain_warning_alone_degrades_the_folded_pair            fixture: this case needs an undegraded node
a_critical_warning_past_the_cap_still_degrades_the_pair    fixture: this case needs an undegraded node
a_malformed_or_flooding_brain_cannot_break_the_nodes_liveness_answer   left: Bool(true) right: false
one_enormous_warning_cannot_bloat_the_liveness_answer      … io stalled: EVERY task in this container was blocked on io 21.7% of the last 10s (some 22.1%) …

The last message is the tell: the node's real degradation probe (PSI io-stall) fired because the GitHub runner was io-stalled, so the "undegraded node" the fixtures start from was already degraded by a true warning about the host. The suite passes 3/3 locally on the same commit (7b06270 main).

This is the same shape as #542 (a test that reads a process-wide surface the code under test does not own): the fixture assumes the degradation registry holds only what the test raised. Options, in order of preference:

1. the fixture clears/snapshots the degradation registry (or disables the host probes) for the duration of a test that asserts on the folded pair;
2. the host-pressure probes carry a source: host marker and the folded-health assertions filter to source: test;
3. skip-with-reason when a host warning is present at test start (weakest — hides the class).

Not fixing in #566 (out of scope); rerunning the job. Filed so the next sighting has somewhere to go.

🤖 Generated with Claude Code