scoreboard: timed OpenEMR arms 3-6, arm 5 API step, agent/screen/API comparison

[abrichr]

Follow-up to #1303, which was merged at 05dd032. These are the scoreboard rows and changes made after that commit. Every figure comes from python -m scoreboard.build over rows committed under packages/flow/benchmark/scoreboard/raw/.

What's in it

• Timed OpenEMR worklist runs: arms 3, 4 and 6, through the throughput harness (feat(flow): parallel worklists with one session per worker, measured #1296, engine b82131e), 26 items each.
  • Arm 3: 8 of 9 correct.
  • Arm 4: 16 of 17 correct.
  • Arm 6: 5 of 8, and 21 of 23 with read-back window 1.
  • Each worklist stops fail-closed at its first uncertain item and pauses the rest; nothing was written at any stop.
  • Causes are filed: OpenEMR: identity check at 'Confirm Create New Patient' reads a background label behind the modal and refuses every 9-17 items #1347 (identity check reads a background label behind the modal), Throughput harness: 'Execution context was destroyed' opening the form in a parallel worker stops the worklist #1348 (a navigation race in the harness's form opening), Throughput harness: synthetic_referrals repeats last names every 26 items, so item 26 trips OpenEMR's duplicate check #1349 (surnames repeat every 26 items).
• Arm 5, the API step (feat(flow): API steps compiled from a recording's network capture #1269, engine 93c32e9): 18 of 18 correct, wall median 3.33 s (n=16 timed). The same bundle with API steps off: 18 of 18, wall median 22.45 s (n=9 timed), interleaved by block on one fixture copy. The possible-duplicate referral r6 is a safe stop in both settings.
• New SCOREBOARD.md section, "OpenEMR: agent, screen program and API step". It is the comparison the README quotes, with each run's profile label, engine, n and load gate.
• Per-run lines under each table: each run's engine, and for worklist runs the harness's own throughput figure. Arm 6 is labelled "self-loaded throughput".
• Imports: an imported run's meta now names the engine that made its rows.
• Protocol: Amendments 4 to 6.
• OpenEMR claims runner (claims_openemr.py): faults are injected only in the fixture database, with triggers and a poller. Its data lands in a later commit, once the run is complete. The build shows a claims run only after its meta is marked complete.

Checks

• ruff check and ruff format --check on benchmark/scoreboard.
• The public artifact inventory was regenerated from a clean checkout of the commit (refresh_inventory.sh).
• Grepped for prospect names; none found.
• Synthetic data only.

🤖 Generated with Claude Code

[abrichr]

Included in batch #1385, which runs CI once for several reviewed PRs. This PR is marked merged when #1385 merges. Please don't push here unless asked; a new commit leaves it out of the batch.