W&B completion seal rejects rewrite run measurements without replace metadata

[lipikaramaswamy]

Summary

W&B publication for rewrite-mode benchmark runs can fail during completion seal validation because the W&B ingress schema requires run.replace even when the measurement record is for mode="rewrite".

Observed error:

ERROR: bad_input error=invalid measurement record on line 1: schema violation at run.replace

This happens after benchmark compute and local analysis complete successfully, so it blocks W&B publication and downstream run comparison.

Context

We hit this while running a RAT Bench rewrite throughput benchmark through the Big Iron/SLURM harness against a PR 206 branch rebased on current main.

The first measurements.jsonl record was a run record with:

{
  "record_type": "run",
  "mode": "rewrite",
  "rewrite": {
    "risk_tolerance": "low",
    "max_repair_iterations": 0,
    "strict_entity_protection": false,
    "has_privacy_goal": true,
    "has_instructions": false
  }
}

No replace metadata is semantically required for a rewrite run, but tools/measurement/measurement_tools/wandb_ingress.py currently defines:

class RunMeasurement(_MeasurementEnvelope):
    ...
    replace: dict[StrictStr, JsonValue]
    rewrite: dict[StrictStr, JsonValue]

So the completion seal rejects rewrite records that do not include replace.

Expected Behavior

Rewrite-mode run records should validate and publish to W&B when they contain rewrite metadata and omit replace metadata, or when replace metadata is explicitly null/empty according to the intended schema contract.

Proposed Fix

Make run metadata schema mode-aware or optional/defaulted for mutually exclusive replace/rewrite config blocks. For example:

• allow replace: dict[...] | None = None
• allow rewrite: dict[...] | None = None
• validate that replace-mode records have replace metadata and rewrite-mode records have rewrite metadata, if strictness is desired
• add a regression test for record_type="run", mode="rewrite" passing write_completion_seal.py

Impact

This blocks W&B publication for rewrite benchmarks even when SLURM/Big Iron execution, scoring, aggregation, and measurement export complete successfully.