Improve v0.9 pack build observability

[AbdelStark]

Summary

Part of #391 under tracker #385.

• add structured progress events for long pass/fail pack builds, persisted as reports/passfail_pack_progress.jsonl and mirrored to stderr with the existing CODELEWM_JOB_EVENT prefix
• extend hf-job-event-status so persisted pack-build logs summarize progress, completion, ETA, and event counts like training job logs
• update the v0.9 short A10G config with the measured cross-benchmark pack p_pass BCE positive-class weight from the preflight pack report
• document the pack-build progress/status workflow in the HF operations and scaled-training runbooks

This PR intentionally does not close #391. The live two-seed HF Jobs run, artifact download, checkpoint inspection, and training health table are still pending after this lands.

Local Validation

• uv run pytest tests/data/execution_pack/test_passfail_pack.py tests/training/test_job_events.py tests/training/test_execution_train_config.py
• uv run python -m compileall -q -x 'tests/fixtures/codestate/invalid_(before|after).py$' codelewm tests
• git diff --check
• uv run pytest tests/ # 967 passed, 8 skipped, 1 warning
• uv run scripts/hf-launch-execution-run --config config/train/scaled/codelewm_execution_v0_9_short_a10g.yaml --git-sha dryrun --date 20260606 --runtime-image-digest sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa --require-runtime-image-digest --json

Pack Preflight Evidence

Built the local v0.9 pass/fail pack with progress logging enabled:

• 384 completion-label rows: HumanEval 282, MBPP-Plus 102
• 2,214 sandboxed scoring inputs
• 2,188 produced records: HumanEval 1,882, MBPP-Plus 306
• pass labels: false 1,107; true 1,081
• p_pass_bce_pos_weight: 1.0240518038852915
• split counts: train 1,928; val 57; test 203
• readiness gates passed, including val/test passed true/false and output_magnitude_bucket coverage
• sandbox rejects: sandbox_timeout 26
• manifest verify ok: 9 files checked
• codelewm secret-scan ok: no findings