Fix main's CI: a Windows-only path bug in my own gate, and folded_health's shared-registry flake

[emooreatx]

main went red at the 0.5.196 merge after eleven consecutive green runs. Two independent failures, one platform each.

This PR carries ci:full deliberately. PRs run ubuntu-only, so the Windows half of this fix cannot be validated by a default PR run — which is exactly why neither failure was catchable before the merge.

Windows — my gate, broken by a path separator

The scrape gate I added in 0.5.196 to prove the node alias literal exists exactly once filtered its own file with a display() string prefix:

.filter(|h| !h.starts_with("src/key_convention.rs"))   // renders src\key_convention.rs on Windows

So on Windows it excluded nothing and the gate failed on the very definition it exists to protect. It now keeps PathBufs and filters with Path::ends_with, which matches whole components and is separator-agnostic. Mutation-verified after the fix: reintroducing a second literal in node_key.rs still reds it.

Ubuntu — one leaked warning, four failures

folded_health shares a process-global degradation registry. Three of the four failures assert degraded_mode == false; the fourth indexes warnings[0], which any node-tier entry displaces. One leaked warning explains all four — and none of the four is the test that leaked it.

I could not reproduce it locally: 10 clean runs unloaded, 8 more pinned to two cores under competing CPU load. So rather than ship a guess, three changes that each stand on their own:

1. tests/common/RaisedWarning — an RAII guard that clears on drop, including while unwinding. The old shape raised, then asserted, bound a loopback port, and made two HTTP round-trips before clearing:

raise("test.forced_node_fault");                   // 575
assert_eq!(baseline, ("degraded", true), ...);     // can panic
spawn_brain(...).await; get_health(...).await;     // can panic
clear("test.forced_node_fault");                   // 593 — skipped on unwind

Correctness no longer depends on the body of a test completing, which is the one thing a test cannot promise. The guard also clears before raising — the defensive shape health_reports_degradation and retention_loop already use and folded_health did not, which is why a leak here was not self-healing.

2. The warnings[0] lookup finds its entry BY CODE. Index zero assumed the brain's warning was first; any node-tier warning ahead of it silently retargeted that assertion at a different entry.

3. The fixture guard PRINTS the registry when it trips. The bare assert!(!baseline.1, "fixture: this case needs an undegraded node") named nothing, so the cause had to be reconstructed from four line numbers in a CI log. Only tests write to that registry, so its contents are the diagnosis:

fixture: `a_brain_warning_alone_degrades_the_folded_pair` needs an undegraded node,
but the process-global degradation registry already says "degraded". Only tests
write to it, so one of them raised without clearing. Standing warnings: [
    Warning { code: "leaked.from_somewhere_else", severity: "critical", ... },
]

Verified by simulating a leak. A recurrence now identifies its own cause.

What I deliberately did not do

health_reports_degradation and retention_loop keep their manual raise/clear pairs. Both already clear defensively before raising, so a leak there self-heals on the next case — the hazard is real but bounded, and churning two passing files to fix a bug they do not have is not obviously right. Happy to convert them if you'd rather have one pattern.

Honest limit

The RAII guard removes a real leak path, but CI showed no panicking test, so I cannot claim it is the cause of that specific run. Changes 2 and 3 are what make the next occurrence cheap: it will either not happen, or it will name itself.

Local

476 lib tests, all 15 gate binaries, and the three degradation-touching suites green. clippy --all-targets -- -D warnings and fmt --check clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_017gxnxsTWS8FkN3CT2JEz86

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0f69e374dd

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Restrict the exclusion to the defining path

Path::ends_with("key_convention.rs") excludes every source file with that basename, not only src/key_convention.rs. If a nested module such as src/foo/key_convention.rs reintroduces the protected literal, this exact-once gate silently passes. Compare against the complete relative path using Path components so the check remains separator-independent without broadening the exclusion.

Useful? React with 👍 / 👎.