Data integrity: score or leaderboard write fails after Finished status

Problem

The compute worker sets submission status to FINISHED before uploading scores and outputs. If push_scores() or push_output() fails (network error, timeout, API failure), the submission is marked Finished but has no scores or outputs.

Root Cause

In compute_worker.py, the flow was:

run.start()         # Sets status to FINISHED at line 1448
if run.is_scoring:
    run.push_scores()  # Upload scores AFTER status update
run.push_output()      # Upload outputs AFTER status update

If push_scores() fails, the submission stays Finished with no scores → data integrity violation.

Impact

• Users see Finished submissions with no results
• Leaderboard is incomplete
• No retry mechanism exists to recover

This bug was discovered during the EEG Foundation Challenge incident analysis.

Solution

1. Reorder operations: upload scores/outputs before setting status to FINISHED
2. Add retry logic to push_scores() with exponential backoff
3. Validate HTTP responses and raise errors on 4xx/5xx

New Flow

run.start()         # Completes scoring but doesn't set FINISHED
if run.is_scoring:
    run.push_scores()       # Upload scores first (with retries)
run.push_output()           # Then upload outputs
if run.is_scoring:
    run._update_status(SubmissionStatus.FINISHED)  # Only now mark as FINISHED

Testing

Added comprehensive K6 integration test:

• tests/k6/test_m8_finished_has_scores.js — verifies invariant: Finished ⟹ has scores
• tests/k6/run_m8_test.sh — bash orchestrator
• Pass criteria: finished_without_scores == 0

Start in compute_worker.py by tracing the run.start(), push_scores(), push_output(), and status-update flow described in the issue. Run tests/k6/run_m8_test.sh and use tests/k6/test_m8_finished_has_scores.js to verify that no Finished submission lacks scores; completion also requires the stated retry and HTTP-response behavior.