Two tests failed intermittently on PR merge-ref lanes without the PRs touching their paths. Each passed on the nextest retry (FLKY-FL 2/2) inside the same job, but the partition still exits non-zero under the fail-on-flake posture. Both look like the same class: a typed authority answered a terminal unavailable inside a mount window instead of a mount-pending answer.
Git topology authority. tracedecay-cli::core_cli_suite cli_non_interactive_test::sessions_git_sync_on_a_cold_daemon_waits_for_the_project_mount panicked at crates/tracedecay-cli/tests/core_cli_suite/cli_non_interactive_test.rs:422 on PR #3074's root-journeys lane (job 111476336231): session git sync did not complete successfully (failed: git_topology_declared_authority_unavailable; remaining work 1). The test holds the daemon after core publish; the sync's topology stage read project_sessions.project_graph_runtime() as None and finished Err(GitTopologySyncFailure::Unavailable) (crates/tracedecay-session-runtime/src/session_sync/git_topology.rs:64) instead of waiting for the bind that lands later in project composition.
External source projection. tracedecay::mcp_suite mcp_handler_test::session_rewrite_test::in_place_edit_of_a_native_id_record_supersedes_every_read_path panicked at crates/tracedecay/tests/mcp_suite/mcp_handler_test/session_rewrite_test.rs:90 on PR #3065's root-transport lane (job 111496854106): tracedecay_hook_runtime returned external_source_projection_unavailable ("Claude transcript ingest failed").
Each needs a root cause rather than a retry. The unbound-runtime early return in publish_git_topology is the first place to check; a project whose session database is read-only never binds it (project_composition.rs), so a mount wait has to distinguish "still mounting" from "never binds".
Two tests failed intermittently on PR merge-ref lanes without the PRs touching their paths. Each passed on the nextest retry (FLKY-FL 2/2) inside the same job, but the partition still exits non-zero under the fail-on-flake posture. Both look like the same class: a typed authority answered a terminal
unavailableinside a mount window instead of a mount-pending answer.Git topology authority.
tracedecay-cli::core_cli_suite cli_non_interactive_test::sessions_git_sync_on_a_cold_daemon_waits_for_the_project_mountpanicked atcrates/tracedecay-cli/tests/core_cli_suite/cli_non_interactive_test.rs:422on PR #3074's root-journeys lane (job 111476336231):session git sync did not complete successfully (failed: git_topology_declared_authority_unavailable; remaining work 1). The test holds the daemon after core publish; the sync's topology stage readproject_sessions.project_graph_runtime()asNoneand finishedErr(GitTopologySyncFailure::Unavailable)(crates/tracedecay-session-runtime/src/session_sync/git_topology.rs:64) instead of waiting for the bind that lands later in project composition.External source projection.
tracedecay::mcp_suite mcp_handler_test::session_rewrite_test::in_place_edit_of_a_native_id_record_supersedes_every_read_pathpanicked atcrates/tracedecay/tests/mcp_suite/mcp_handler_test/session_rewrite_test.rs:90on PR #3065's root-transport lane (job 111496854106):tracedecay_hook_runtimereturnedexternal_source_projection_unavailable("Claude transcript ingest failed").Each needs a root cause rather than a retry. The unbound-runtime early return in
publish_git_topologyis the first place to check; a project whose session database is read-only never binds it (project_composition.rs), so a mount wait has to distinguish "still mounting" from "never binds".