B-Y1NS-quality-gate-bun-test-dashboard-server-races-as-spurious-new
Status: closed/abandoned
Surfaced by T-0008 (PR #356, rumdl wiring). The bun test quality-check
verb non-deterministically reports new-drift even on a markdown-only
branch: a dashboard test races a transient HTTP server up and down, so
sdlc quality run --diff-against-baseline flags 2 findings whose pid and
port differ every run (e.g. pid 10230 :56521, pid 5974 :52586). The
baseline-diff gate cannot filter a finding that is randomly present —
whether the server is caught depends on test timing, not on the branch.
Effect: every task-work run whose gate includes bun test may show a
spurious FAIL new-drift=2, forcing the operator to manually confirm
it is the known flaky rather than real drift. T-0008 hit this and had to
prove the branch touched no server code to clear it.
Two candidate fixes (triage to pick):
- Make the dashboard test deterministic — bind to a fixed sentinel or
tear the server down synchronously before the test returns, so it is
never observable to a concurrent
quality run. - Teach the gate’s
bun testfinding-parser to ignore transient server-listing lines (http://127.0.0.1:<port> (pid <n>)and the<TMPDIR> <pid> <url>shape), so ephemeral servers never count as drift.
Found-in: T-0008 / D-0011-markdown-formatting-tool post-mortem.
Closed 2026-08-16: duplicate of T-NRTX-mask-bare-pid-columns-in-findings and T-B75W-dashboard-test-isolates-live-dashboards, which carry the two candidate fixes named here.