HOMEGOLF-19: Build the retained local-cluster HOMEGOLF scoreboard, delta ledger, and stop-condition loop

[AtlantisPleb]

Summary

Turn the local clustered HOMEGOLF lane into a retained score-improvement loop with explicit deltas, promotion rules, and a stop condition.

Tracking issue: #633

Why

The current local HOMEGOLF loop has real runs, but not yet one explicit operator truth for:

• which run is the incumbent
• which run improved the incumbent
• which changes were system-only versus model-side
• which failures are blockers versus ordinary regressions
• when to stop iterating locally and promote a candidate upward

The user goal is to keep improving until there are no meaningful improvements left. That requires a scoreboard and stop condition, not only audits.

Scope

• define one canonical retained local-cluster HOMEGOLF scoreboard or ledger
• store per-run config, artifact paths, score receipts, and delta versus current incumbent
• classify each run as improvement, noise, regression, blocked, or invalid
• distinguish score changes caused by system or runtime fixes from score changes caused by model-family changes
• define the local stop condition for “no more meaningful improvements observed”
• tie promotion gating to that same ledger instead of ad hoc judgment

Acceptance Criteria

• every retained local HOMEGOLF candidate run has one machine-legible score row
• the loop can identify the incumbent and the exact delta from each challenger
• the scoreboard can say honestly when local improvements have plateaued
• promotion to H100 uses the same scoreboard instead of a separate manual story

References

• docs/HOMEGOLF_TRACK.md
• docs/2026-03-28-parameter-golf-winner-gap-and-psionic-path-audit.md

[AtlantisPleb]

This issue is superseded by the current lane docs and retained operator surfaces.

Since it was opened, the repo has landed and retained:

• daily scoreboard surfaces
• HOMEGOLF public-comparison and competitive-ablation reports
• canonical HOMEGOLF track truth in docs/HOMEGOLF_TRACK.md

That means the original need for "one scoreboard instead of only audits" has already been partly absorbed into shipped lane surfaces. The remaining stop-condition and promotion-loop work is still real, but it is not the active repo critical path now.

So I am not keeping this older tactical scoreboard issue open in parallel with the newer canonical lane docs. If HOMEGOLF score-loop closure becomes active again, reopen it as a fresh narrower issue tied to the current retained scoreboard contract.

Closing as superseded / not planned.