Carry the rollout id on both agentic paths

[fzyzcjy]

What breaks

• tests/e2e/sglang/test_session_v1_v2_parity.py::test_qwen3_8b_h200_fa3_agentic_v2_drop_retries_matches_v1_training_payload_bitwise fails on main:

AssertionError: sample.rollout_id is not bitwise equal
left = None, right = 0, path = 'sample.rollout_id'

• It is not flaky and not environment-dependent. The scheduled run on main reproduces it: 31715141995 → stage-c-2-gpu-h200 (0), Test Summary: 4/5 passed.

Root cause

• miles/rollout/generate_hub/agentic_tool_call.py stamps the rollout id only on the v2 branch:

    samples = result.samples
    if use_v2:
        rollout_id = input.sample.rollout_id if input.sample.rollout_id is not None else input.sample.index
        assert rollout_id is not None, "v2 agentic samples require input Sample.rollout_id or Sample.index"
        for sample in samples:
            sample.rollout_id = rollout_id

• So v1 leaves sample.rollout_id at None while v2 sets it to the index. The parity test compares every field of the training payload, so the two paths are structurally unable to agree on this one.
• The two halves landed independently and were never reconciled:

Change	Date	Effect
#2129 test(session): validate H200 v1/v2 agentic parity	2026-08-05	Adds the bitwise parity test over all payload fields
#2368 fix(rollout): group session v2 leaf samples	2026-08-11	Adds the v2-only sample.rollout_id stamping

• The rollout id is sample identity, not a v2 concept — nothing about it is specific to the session-server path.

The fix

• Hoist the derivation out of the branch, keep the assertion where it was, and stamp whenever a value is derivable:

    samples = result.samples
    rollout_id = input.sample.rollout_id if input.sample.rollout_id is not None else input.sample.index
    if use_v2:
        assert rollout_id is not None, "v2 agentic samples require input Sample.rollout_id or Sample.index"
    if rollout_id is not None:
        for sample in samples:
            sample.rollout_id = rollout_id

• v2 is byte-identical in behaviour: the same assertion still guards only that path, and rollout_id is non-None there by construction, so the new if is always taken.
• v1 gains the stamp only when derivable: if neither rollout_id nor index is set, v1 keeps its current behaviour rather than acquiring a new failure mode.

Evidence

• The same change, applied on a chain based on 6033ac0eaf, turns that lane green: stage-c-2-gpu-h200 (0) went from Enabled 5 test(s) → 4/5 passed to Enabled 5 test(s) → 5/5 passed, running 51 minutes, so it is a real pass rather than a skipped or no-op lane.
• ruff check miles/rollout/generate_hub/agentic_tool_call.py passes.

Note for reviewers

• [AMD] Enable amd pr ci #2347 edits the same guard to if use_v2 and len(samples) > 1: for an unrelated reason. Whichever lands second will need a trivial rebase; the two intents are compatible, since the condition being narrowed there is the stamping loop rather than the derivation.

[guapisolo]

Rebased fix direction after #2347: this branch now keeps rollout_id=None for v1 and v2 singleton results, and assigns a shared rollout ID only to v2 multi-leaf results via if use_v2 and len(samples) > 1.

This preserves ordinary postprocessing, including sample-count trimming and dynamic-global-batch metadata, while retaining v2 leaf grouping. Focused verification passed: tests/fast/rollout/generate_hub/test_agentic_v2.py (7 passed) and Ruff on agentic_tool_call.py.

The PR description still describes the superseded v1-stamping approach; this comment records the final behavior instead.

[guapisolo]

Do changes to align with session server v2 design