fix(toml): quote string values in the generated Hydra overrides

[kevin9327]

Problem

A string value in an SFT TOML can be changed or rejected on its way into the config. _hydra_format only quotes a top-level string when it holds a comma or a space, and Hydra re-types an unquoted value:

TOML value (string field)	Override emitted	What the config receives
name = "20260913"	job.name=20260913	int 20260913
name = "1e-4"	job.name=1e-4	float 0.0001
name = "true" / "null"	job.name=true / job.name=null	True / None
name = "lr=1e-5"	job.name=lr=1e-5	OverrideParseException: mismatched input '=' expecting <EOF>
name = "run(v2)"	job.name=run(v2)	Unknown function 'run'
name = "run[1]", "exp#3"	unquoted	parse error
name = "bob's run"	job.name='bob's run'	parse error (quote not escaped)

So a run name, group, project or path containing =, (, [, # or an apostrophe stops --sft-toml loading before training starts. A numeric-looking run name loads but is an int, which Config.validate then passes to bytearray(self.job.name, "utf-8"), and that raises TypeError: encoding without a string argument. String items inside lists were already quoted, but their apostrophes weren't escaped either.

Fix

Quote every string value, in lists or not. The new _hydra_quote escapes ' and doubles only the backslashes Hydra would otherwise consume, meaning those just before a quote or at the end of the string. Numbers, bools, lists, ${oc.env:...} interpolation (still resolved inside quotes) and the "???" skip are unchanged. The now-unused in_list parameter is removed.

Tests

New tests in cosmos_framework/configs/toml_config/sft_config_test.py (TestBuildHydraOverrides) feed the emitted override back through Hydra's own OverridesParser:

• test_string_value_reaches_hydra_as_the_same_string: parametrized over the values in the table plus a\'b and ends\. The result must be a str equal to the TOML value.
• test_string_list_items_reach_hydra_as_the_same_strings: ["480", "bob's"].
• test_non_string_values_keep_their_types: lr=1e-5, betas=[0.9, 0.99], max_iter=200, ema.enabled=false, vae_path="${oc.env:WAN_VAE_PATH}" and load_path="???" parse to the same types and values as before. This test passes both before and after the change, so the change doesn't widen what is emitted for non-strings.

Fail-before (source reverted, tests kept):

E   AssertionError: job.name=20260913
    assert (False)
     +  where False = isinstance(20260913, str)
E   AssertionError: job.name=true
    assert (False)
     +  where False = isinstance(True, str)
E   hydra.errors.HydraException: mismatched input '=' expecting <EOF>
E   hydra.errors.HydraException: HydraException while evaluating 'run(v2)': Unknown function 'run'
E   hydra.errors.HydraException: token recognition error at: '#'
E   hydra.errors.HydraException: mismatched input 's' expecting <EOF>
E   hydra.errors.HydraException: token recognition error at: '']'
FAILED .../sft_config_test.py::TestBuildHydraOverrides::test_string_value_reaches_hydra_as_the_same_string[20260913]
FAILED .../sft_config_test.py::TestBuildHydraOverrides::test_string_value_reaches_hydra_as_the_same_string[lr=1e-5]
FAILED .../sft_config_test.py::TestBuildHydraOverrides::test_string_value_reaches_hydra_as_the_same_string[bob's run]
FAILED .../sft_config_test.py::TestBuildHydraOverrides::test_string_list_items_reach_hydra_as_the_same_strings
11 failed, 23 passed, 2 skipped

(ends\ passes before too: an unquoted trailing backslash already parsed.)

Counts for sft_config_test.py:

	passed	failed	skipped
main, existing tests	21	0	2
new tests, source reverted	23	11	2
new tests, with fix	34	0	2

The shipped examples/toml/sft_config/*.toml still validate and build overrides; those parametrized cases are part of the 34.

Run on CPU: PYTHONUTF8=1 COSMOS_TRAINING=0 pytest cosmos_framework/configs/toml_config/sft_config_test.py. On a non-UTF-8 locale, PYTHONUTF8=1 is needed for the existing example-TOML test to read the files. COSMOS_TRAINING=0 avoids the object-store backend import.

Lint

ruff 0.12.7 (repo pin) on both touched files: ruff check output is identical before and after (one pre-existing I001 in toml_config_helper.py). ruff format --diff is the same size before and after (28 changed lines, all pre-existing). sft_config_test.py is clean both ways.

🤖 Generated with Claude Code