fix: dynamic var spec serialization, JSON indexes, and StepMutator sequencing

[npow]

Summary

Three bug fixes for the dynamic decorator values framework from #2970, found during integration testing with the Maestro orchestrator.

1. Preserve dynamic var spec metadata (df8dc15f)

Decorator.translate_decorator_spec was reconstructing DynamicVar objects from the __dynvar__: sentinel but losing pertask and default metadata. Fixed by base64-encoding a JSON payload into the sentinel so all fields survive the spec → string → spec roundtrip.

2. Handle JSON stringified dynamic var indexes (d55d75e4)

Foreach split indexes stored as Python int keys ({0: "val"}) become string keys ({"0": "val"}) when serialized through JSON (e.g., via Maestro). resolve_dynamic_vars_from_store now accepts both integer and stringified integer keys.

3. Fix StepMutator pre-mutate sequencing (2bc95e1e)

FlowSpec._pre_step_decorator_hooks was calling step_decorator.pre_mutate_step for every step on every step_decorator, but StepMutator.pre_mutate_step should only be called on the step that owns the mutator. Fixes incorrect dynamic var resolution when multiple steps have mutators.

PR Type

• Bug fix

Test plan

• Unit tests added for each fix (test/unit/test_dynamic_var_decorator_spec.py)
• Integration tested against Netflix Maestro with foreach, switch, and task-process decorator patterns

Fixes bugs in #2970.

[greptile-apps]

Greptile Summary

This PR fixes several dynamic decorator value roundtrip issues. The main changes are:

• Dynamic var decorator specs now carry pertask and default metadata.
• Foreach dynamic var lookup now accepts JSON-stringified split indexes.
• StepMutator pre-mutation now runs on the mutator-owning step path.
• Unit tests cover the new decorator spec and index behavior.

Confidence Score: 4/5

The dynamic-var spec serialization path needs fixes before merging.

• Non-JSON Python defaults can now fail during task launch or deployment.
• Legacy dynamic var names beginning with json: can be parsed as the new encoded format and raise instead of roundtripping.
• The split-index and StepMutator sequencing changes otherwise match the surrounding contracts.

metaflow/decorators.py

Important Files Changed

Filename	Overview
metaflow/decorators.py	Adds JSON/base64 dynamic-var sentinel encoding and decoding, but the new format can crash on non-JSON defaults and misread legacy names beginning with json:.
metaflow/dynamic_var.py	Adds stringified split-index fallback for pertask dict resolution while preserving integer-key precedence.
metaflow/flowspec.py	Runs StepMutator pre-mutation over config_decorators and removes the duplicate pre-mutate pass.
test/unit/test_dynamic_var_decorator_spec.py	Adds tests for legacy sentinel parsing, metadata preservation, nested dynamic vars, and stringified split indexes.

_{Reviews (1): Last reviewed commit: "Fix StepMutator pre-mutate sequencing fo..." | Re-trigger Greptile}

[greptile-apps]

Default Serialization Can Crash

When a decorator dynamic var uses a valid Python default such as datetime, bytes, an enum, or a custom object, this path stores that object in the JSON payload and json.dumps(payload) raises during make_decorator_spec(). The resolver accepts defaults as Python values and returns them directly, so preserving the metadata this way can turn task launch or deployment into a serialization failure.

[greptile-apps]

Legacy Json Names Misdecode

An existing legacy sentinel for a variable whose name starts with json:, such as __dynvar__:json:foo, now enters the new base64 decode branch instead of the legacy fallback. That makes parsing the decorator spec raise from urlsafe_b64decode rather than reconstructing DynamicVar("json:foo"), breaking backward compatibility for those saved specs.