feat(sc): run the distillation teacher in the SingleController train pump

[tianyi-zhang-02]

What does this PR do?

Wires distillation into the SingleController train pump. It builds and parks a teacher TQPolicy, scores each chunk, writes the teacher top-k tensors to TransferQueue, narrows the student train fields, and skips the reward/advantage work that distillation does not use.

The same change also closes the setup gaps exposed by the first real run: teacher schema registration and shutdown, teacher train_iters, tokenizer compatibility, distillation-safe algorithm validation, and the missing setup timing.

This is the middle PR in the #3843 → #3846 → #3849 stack. The recipe and checkpoint/restore functional test stay in #3849.

Ownership

The stack is on hold because its core teacher top-k entry point overlaps the still-open #2580. I am keeping this branch tested and current, but not extending it until that ownership question is resolved.

Validation

• Final branch b8073a3d894bbad0f7c5a6c67c4faa92dfbe795e is aligned with upstream main at ccbcd4cc5442784f6af2288dd99021560480b8f2.
• The full stacked unit/config sweep at test(sc): add the distillation recipe and functional test for SingleController #3849 head 907606048aa8c950027dbc199bff1495035e862c passed 765 cases with 6 GPU-only skips.
• The stacked test(sc): add the distillation recipe and functional test for SingleController #3849 functional test passes on a 4×H100 80 GB host (2 GPUs used): steps 1–2 train and checkpoint, run 2 restores the native data-plane state and trains steps 3–4.
• Both runs report finite non-negative loss below the test bound and include teacher initialization timing.
• NCCL_NVLS_ENABLE=0 was set for this Runpod H100 topology, matching the workaround already used by H100 recipes in this repository.

[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.

[tianyi-zhang-02]

Heads-up on merge order, with the resolution written out.

Four of my open PRs edit the same twelve-line block — the unsupported knob list in _validate_algo_settings:

• fix(sc): preserve DatumSpec loss_multiplier through rollouts #3786 removes overlong_filtering from it (build_sample_mask now applies it)
• fix(sc): apply advantage clipping and reward scaling in the advantage stage #3787 removes reward_scaling from it (_advantage_stage now applies it)
• feat(sc): run the distillation teacher in the SingleController train pump #3846 / test(sc): add the distillation recipe and functional test for SingleController #3849 wrap the list in an is_distillation_run guard, because DistillationConfig defines none of those knobs and the comprehension would AttributeError

I checked every pair by actually merging them. #3786 × #3787 is clean now — they only ever conflicted on the comment wording and on both adding a test before the same anchor, and both are fixed. The remaining four pairs are a genuine textual conflict on that block, and no restructuring avoids it: extracting or reordering the list conflicts just as hard.

So rather than force a stack, here is the resolution. Whichever order they land in, this is the merged form:

    # An enabled one here describes shaping this run does not do. An entry
    # leaves this list when the SC path starts implementing it -- rejecting a
    # knob is only right while nobody honours it.
    #
    # DistillationConfig defines none of them: there is no reward to shape or
    # filter on, so the list cannot even be evaluated on that path.
    unsupported = (
        []
        if is_distillation_run(master_config)
        else [
            name
            for name, enabled in (
                ("use_dynamic_sampling", algo_cfg.use_dynamic_sampling),
                ("reward_scaling", algo_cfg.reward_scaling.enabled),
                ("reward_shaping", algo_cfg.reward_shaping.enabled),
            )
            if enabled
        ]
    )

I built that merge locally and ran the SC and config suites against it — 1084 passed. Happy to rebase whichever ones are left once the first lands; just say which order you want.