HOMEGOLF-15: Drive the local-cluster HOMEGOLF score-improvement loop to operational closure

[AtlantisPleb]

Summary

Use the reachable local clustered HOMEGOLF devices as the primary proving ground for the Rust-native Parameter Golf lane until system changes are no longer blocked by operational failures and are visibly improving retained local scores.

This is the main tracking issue for the next loop.

Why

Current retained reality:

• Psionic's retained real full-validation PGOLF score is still 6.306931747817168
• recent local clustered HOMEGOLF runs exposed real operational blockers instead of a clean score-improvement loop
• 20260328f, 20260328k, and 20260328l exported artifacts but still did not close into detached artifact_score_report.json receipts
• 20260328j died with cudaMalloc failed: out of memory
• 20260328m hit NaN on step 2 and then panicked during EMA final-surface application

That means the immediate job is not “turn the big GPUs back on and hope.” The immediate job is to drive out Rust runtime, score-closeout, and operator-kink failures on the local clustered HOMEGOLF lane until system changes produce retained score movement.

Intent

The operating rule from here is:

• local clustered HOMEGOLF devices first for Rust/runtime/ops closure
• require retained local score improvement from those system changes
• promote only de-risked candidates to H100
• use 8xH100 only after the current Rust lane stops tripping over avoidable blockers

Scope

This master task tracks the issue stack required to enter an honest improvement loop:

• fix detached score-closeout so artifact-only runs actually produce score receipts
• stabilize the local clustered competitive runner under explicit grad-accum, raw, and EMA postures
• align the exact competitive lane to the public 11L winner family where Psionic is still behind
• build a retained local-cluster scoreboard and stop-condition loop
• define the single-H100 and later 8xH100 promotion gate
• prove whether XTRAIN improves the score path or keep it off the critical path

Acceptance Criteria

• the subordinate issue stack is closed
• the local clustered HOMEGOLF lane can run without the current score-closeout, OOM, NaN, or queue-truth failures
• at least one retained local clustered HOMEGOLF score receipt improves because of concrete Rust-system or model-lane changes
• the current best local clustered candidate has an explicit promotion verdict for single-H100 and later 8xH100

References

• docs/2026-03-28-parameter-golf-winner-gap-and-psionic-path-audit.md
• docs/2026-03-28-homegolf-xtrain-pgolf-explicit-grad-accum-queue-correction-audit.md
• docs/HOMEGOLF_TRACK.md

[AtlantisPleb]

Issue stack created for #633:

• HOMEGOLF-16: Unblock detached artifact scoring and score-closeout receipts on the local HOMEGOLF lane #634 HOMEGOLF-16: Unblock detached artifact scoring and score-closeout receipts on the local HOMEGOLF lane
• HOMEGOLF-17: Stabilize the local clustered competitive HOMEGOLF runner under explicit grad-accum and EMA surfaces #635 HOMEGOLF-17: Stabilize the local clustered competitive HOMEGOLF runner under explicit grad-accum and EMA surfaces
• HOMEGOLF-18: Align the exact HOMEGOLF competitive lane with the public 11L winner family #636 HOMEGOLF-18: Align the exact HOMEGOLF competitive lane with the public 11L winner family
• HOMEGOLF-19: Build the retained local-cluster HOMEGOLF scoreboard, delta ledger, and stop-condition loop #637 HOMEGOLF-19: Build the retained local-cluster HOMEGOLF scoreboard, delta ledger, and stop-condition loop
• HOMEGOLF-20: Promote only de-risked HOMEGOLF candidates from the local cluster to H100 and later 8xH100 #638 HOMEGOLF-20: Promote only de-risked HOMEGOLF candidates from the local cluster to H100 and later 8xH100
• XTRAIN-24: Prove HOMEGOLF score gain from XTRAIN or keep XTRAIN off the critical path #639 XTRAIN-24: Prove HOMEGOLF score gain from XTRAIN or keep XTRAIN off the critical path

Recommended execution order:

1. HOMEGOLF-16: Unblock detached artifact scoring and score-closeout receipts on the local HOMEGOLF lane #634
2. HOMEGOLF-17: Stabilize the local clustered competitive HOMEGOLF runner under explicit grad-accum and EMA surfaces #635
3. HOMEGOLF-19: Build the retained local-cluster HOMEGOLF scoreboard, delta ledger, and stop-condition loop #637
4. HOMEGOLF-18: Align the exact HOMEGOLF competitive lane with the public 11L winner family #636
5. XTRAIN-24: Prove HOMEGOLF score gain from XTRAIN or keep XTRAIN off the critical path #639
6. HOMEGOLF-20: Promote only de-risked HOMEGOLF candidates from the local cluster to H100 and later 8xH100 #638

[AtlantisPleb]

This March 28 HOMEGOLF umbrella is no longer the right active tracker.

What changed since it was opened:

• docs/HOMEGOLF_TRACK.md now owns the canonical HOMEGOLF lane contract and retained surfaces
• docs/ROADMAP_PARAMETERGOLF.md owns the Parameter Golf lane-specific roadmap and stop record
• the repo now has the unified PSION-* executor backlog (#698 through #787), and that is the current critical path
• the subordinate March 28 tactical issues under this umbrella have now been retired accordingly

The important program decision is:

• keep HOMEGOLF as an active local-first proving lane
• do not keep this older tactical closure umbrella open in parallel with the newer lane authorities and active PSION-* stack

If HOMEGOLF becomes the active bottleneck again, reopen fresh scoped HOMEGOLF-* issues against the current lane contract rather than reviving this pre-roadmap umbrella.

Closing as superseded / not planned.