Workflow.to_yaml() emits unquoted strings like 1e-5 that Argo/kubectl parse as numbers #1607

Pre-bug-report checklist

1. This bug can be reproduced using pure Argo YAML

• Yes 👉 Please report a bug to the Argo Workflows GitHub 👈
• No

If yes, it is more likely to be an Argo bug unrelated to Hera. Please double check before submitting an issue to Hera.

2. I have searched for existing issues

• Yes (closest prior work: Workflow.to_yaml() emits unquoted y/Y/n/N` strings that Argo/kubectl parse as booleans #1597 / Fix YAML dump quoting for y/Y/n/N boolean aliases #1598, which fixed the y/Y/n/N boolean case in the same representer)

3. This bug occurs in Hera when...

• exporting to YAML
• submitting to Argo
• running on Argo with the Hera runner
• other:

Bug report

Describe the bug

Workflow.to_yaml() / hera._yaml.dump() emit some string values unquoted that go-yaml (used by Argo and kubectl) parses as numbers. PyYAML decides whether to quote a string with its own YAML 1.1 resolver, but go-yaml resolves more plain scalars as numbers, so anything in the gap is written unquoted and changes type on the way to the cluster.

The gap, checked against gopkg.in/yaml.v2 and sigs.k8s.io/yaml (the path kubectl apply takes):

PyYAML already quotes the forms its own resolver knows (1.5E-3, .5, 5., 0x1F, 0b11, 1_000, 010, .inf), so this only affects the forms above. The realistic hit is a parameter or env value such as a learning rate "1e-5" or an epoch count "1e3".

Error log if applicable:

With the Argo CRDs from install.yaml (they carry a schema), kubectl apply rejects the exported YAML:

The WorkflowTemplate "yaml-number-repro" is invalid:
* spec.arguments.parameters[0].value: Invalid value: "number": spec.arguments.parameters[0].value in body must be of type string: "number"
* spec.arguments.parameters[1].value: Invalid value: "integer": spec.arguments.parameters[1].value in body must be of type string: "integer"
* spec.templates[0].container.env[0].value: Invalid value: "number": spec.templates[0].container.env[0].value in body must be of type string: "number"

With the minimal CRDs (no schema) the apply succeeds and the values are stored with the wrong type: learning-rate becomes the float 1e-05, epochs becomes the int 1000, LEARNING_RATE becomes the float 1e-05.

To Reproduce

Full Hera code to reproduce the bug:

from hera.workflows import Container, Env, Parameter, WorkflowTemplate

with WorkflowTemplate(
    name="yaml-number-repro",
    entrypoint="train",
    arguments=[Parameter(name="learning-rate", value="1e-5"), Parameter(name="epochs", value="1e3")],
) as wt:
    Container(
        name="train",
        image="python:3.12-alpine",
        command=["python", "-c", "print('ok')"],
        env=[Env(name="LEARNING_RATE", value="1e-5")],
    )

print(wt.to_yaml())

Output (values unquoted):

      env:
      - name: LEARNING_RATE
        value: 1e-5
  arguments:
    parameters:
    - name: learning-rate
      value: 1e-5
    - name: epochs
      value: 1e3

Then kubectl apply -f on that output gives the error above (Argo v4.1.2 CRDs on a kind cluster).

Expected behavior

Strings that go-yaml would read as a number should be quoted, the same way y/n are quoted since #1598:

      env:
      - name: LEARNING_RATE
        value: '1e-5'
  arguments:
    parameters:
    - name: learning-rate
      value: '1e-5'
    - name: epochs
      value: '1e3'

With that output kubectl apply succeeds and the values are stored as strings.

Environment

• Hera Version: main (ee518f5, after 7.1.0)
• Python Version: 3.14.4 (PyYAML 6.0.3)
• Argo Version: 4.1.2 CRDs, kubectl 1.37

Additional context

go-yaml resolves a plain scalar as a number when, after removing underscores, strconv.ParseInt(s, 0, 64) succeeds or the scalar matches its yamlStyleFloat regex ^[-+]?(\.[0-9]+|[0-9]+(\.[0-9]*)?)([eE][-+]?[0-9]+)?$ (see resolve.go in gopkg.in/yaml.v2). The str_presenter in src/hera/_yaml.py can mirror those two rules and single-quote matching strings. Quoting a string is always safe for Argo, so erring on the side of quoting does not change any value. I have a PR ready with the fix and tests.