Why: A timed window that starts under load average above 16 is not a usable baseline, and creating the instance first only to pause wastes a startup.
Evidence: The rerun at load 24.03 exited 3 before any timed window and removed the instance.
Impact: The pause now happens before initdb. No change to the benchmark itself.
Refs: docs/release-multi-instance.md B6
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The percentile reader also opened the pgbench stdout file, so the first run exited 1 after six successful windows.
Evidence: Parser check on results/pgbench-20261012-131229 select-only-10 gave p50 0.018 ms, matching the reported average. The timed rerun is recorded next.
Impact: Percentile parsing only. The six windows themselves had zero failed transactions.
Refs: docs/release-multi-instance.md B6
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: backend U4 password gate returns 403 on business APIs until must_change clears; CLI-created load passwords are terminal.
Evidence: dual smoke failed POST /api/projects 403 after login_keys pin.
Impact: load suite works on schema v26+; EXIT trap disables re-entry and waits for serve.
Refs: server enforcePasswordGate; CreateAccount must_change
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: background subshell PID was recorded instead of serve, so cleanup left orphans that stole ports and broke the next smoke.
Evidence: prior dual evidence failed with bind address already in use + probe 401 against stale instance.
Impact: trap kill now terminates the real serve process.
Refs: tests/load/collaboration/run_load.sh
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: collaboration storage U7 (proposal sections 3/5/7) requires two-process
HTTP-level acceptance over shared PG, a reproducible load-test entry honoring
the load contract, and an auditable pre-release checklist; capacity windows and
dual-host runs need user-scheduled resources, so the deliverable here is the
verifiable entry, not capacity claims.
Evidence: make test green (composite); make test-multi-instance green locally
against an isolated PG (five scenarios incl. crash-takeover and restart
invariants); run_load.sh --profile smoke end-to-end green (SSE probe p95~18ms,
zero seq duplicates, stack auto-recycled); backend pinned to feature branch
with the three cross-instance fixes.
Impact: new tests/integration/multi-instance (CI test-pg job step), new
tests/load/collaboration (k6 + SSE probe + run_load.sh), release checklist at
docs/release-multi-instance.md; no runtime behavior change in Dev repo itself.
Refs: backend PR #164; task card feature/u7-load-fault-release.
agent-exec-transport
weichenandCursor