[iris] get_job_info crashes on constraint JSON with mode field

[dlwh]

Describe the bug
iris.cluster.client.get_job_info() crashes when IRIS_JOB_CONSTRAINTS contains a constraint JSON object with a mode field. This breaks executor-managed child jobs before user code runs. In this thread it caused the baseline arm of resilient-1e18-0325a-dpequiv to fail before training started.

To Reproduce

1. Launch an executor-managed TPU step that calls iris.cluster.client.get_job_info() during startup, for example run_elastic_budget_compare_baseline_direct in lib/marin/src/marin/training/elastic_budget_compare.py:326.
2. Ensure the inherited IRIS_JOB_CONSTRAINTS environment contains a constraint JSON object with a mode field.
3. Observe startup fail before training begins.

Expected behavior

get_job_info() should tolerate the current serialized constraint shape and return JobInfo instead of crashing on protobuf JSON parsing.

Additional context

Failure is in lib/iris/src/iris/cluster/client/job_info.py:101:

google.protobuf.json_format.ParseError: Message type "iris.cluster.Constraint" has no field named "mode" at "Constraint".
Available Fields(except extensions): "['key', 'op', 'value', 'values']"

The current code still does json_format.ParseDict(item, cluster_pb2.Constraint()), so older/newer serialized forms of IRIS_JOB_CONSTRAINTS are not schema-tolerant. A minimal fix is to normalize or ignore unknown fields when reconstructing inherited constraints in get_job_info().

[github-actions]

🤖 I'm looking into this issue. I'll post a task list shortly with my plan and progress.

[dlwh]

🤖 # Research
get_job_info() still reconstructs inherited constraints by parsing IRIS_JOB_CONSTRAINTS with json_format.ParseDict(item, cluster_pb2.Constraint()). That breaks when the serialized constraint JSON includes a mode field, so executor-managed child jobs fail before user code runs. The failure is triggered from the baseline path in lib/marin/src/marin/training/elastic_budget_compare.py:326.

Relevant code

• lib/iris/src/iris/cluster/client/job_info.py:101 - constraint parsing
• lib/marin/src/marin/training/elastic_budget_compare.py:326 - startup path that calls get_job_info()
• lib/iris/src/iris/cluster/runtime/env.py:105 - emits IRIS_JOB_CONSTRAINTS

Proposed Fix

Make the constraint parse schema-tolerant in get_job_info() by ignoring unknown fields when converting JSON to cluster_pb2.Constraint, or by normalizing the inherited constraint payload before parsing. Add a regression test that sets IRIS_JOB_CONSTRAINTS to a JSON object containing mode and verifies get_job_info() still returns JobInfo.

[claude]

🤖 Plan:

• Research codebase and identify root cause
• Post research + proposed fix comment
• Create branch and implement fix with test
• Run pre-commit and tests
• Open PR
• Verify CI

[claude]

🤖 Plan:

• Research codebase and identify root cause
• Post research + proposed fix comment
• Create branch and implement fix with test
• Run pre-commit and tests
• Open PR: [iris] Fix get_job_info crash on constraint JSON with unknown fields #4169
• Verify CI (lint-and-format ✅, iris-tests ✅)