fix(ci): use ++ for token_capture override in SC functional tests

[terrykong]

What does this PR do ?

Unblocks L1_Functional_Tests_SingleController (and its GB200 twin) on #4022.

Both jobs fail before training starts:

omegaconf.errors.ConfigKeyError: Key 'token_capture' is not in struct
hydra.errors.ConfigCompositionException: Could not override 'token_capture.enabled'.
nemo_rl.utils.config.OverridesError: Failed to parse Hydra overrides: Could not override 'token_capture.enabled'.

L1_Functional_Tests_SingleController.sh passes token_capture.enabled=true as a plain Hydra
override, but the config that test runs — examples/nemo_gym/grpo_qwen3_30ba3b_instruct.yaml —
has no token_capture block. Hydra composes in struct mode, so overriding a key that is not in
the YAML is an error. MasterConfig.token_capture having a Pydantic default does not help,
because that default is applied after Hydra composition.

The fix is the ++ (add-or-override) prefix, which is how this same script already passes the
other SingleController-only knobs into this shared Gym config — ++data_plane.*,
++async_rl.*. One line.

Details

This is not a #4022 regression. It reproduces on main:

• main's L1_Functional_Tests_SingleController.sh has the same un-prefixed override.
• main's grpo_qwen3_30ba3b_instruct.yaml has no token_capture key.
• ci: Bump Megatron-Bridge to 5ed9799  #4022 touches neither file.

It came in with #3837 (7036e5d16), which added the test line and the token_capture feature but
did not add the key to the Gym config the test runs. Targeting #4022's branch because that is what
the failure is currently blocking; happy to retarget at main instead if you would rather it land
independently.

Why not add token_capture: to the YAML instead? That config is shared by eight other
functional scripts, most of them not SingleController, so declaring an SC-only block there has a
wider blast radius than the caller-side ++. Either fix works — say the word if you prefer the
config-side one and I will switch it.

The test still asserts what it is meant to. grpo_async_gym_single_controller.sh gates its
extra check on the argument string:

if [[ "$*" == *token_capture.enabled=true* ]]; then
    EXTRA_CHECKS+=('max(data["train/finalize/total_ms"]) > 0')
fi

++token_capture.enabled=true still matches that glob, so the finalizer assertion keeps running —
the test is not being quietly turned into a no-op.

Verification

Reproduced and fixed locally, CPU only, against the real config loader:

token_capture.enabled=true    -> FAILED: Could not override 'token_capture.enabled'.
++token_capture.enabled=true  -> OK, token_capture.enabled = True

(driving nemo_rl.utils.config.load_config + parse_hydra_overrides on
examples/nemo_gym/grpo_qwen3_30ba3b_instruct.yaml, which is the exact path that raises in CI)

The glob check above was verified separately in bash. The GPU run itself is not something I can
execute — that needs CI.

Issues

Unblocks CI on #4022.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[terrykong]

Closing — @cspades landed the same fix independently in 28f4d08. No longer needed.