feat(campaign): runLineageLoop — Lineage DAG live seams

[drewstone]

Wires the just-merged Lineage DAG primitive (src/campaign/lineage.ts — Lineage, runLineage, Governor, heuristicGovernor) to the real improvement machinery so the multi-track, multi-parent improvement DAG can run live.

What this adds

src/campaign/presets/run-lineage-loop.ts — runLineageLoop, a thin additive adapter that supplies runLineage's two abstract seams:

• step = ONE SurfaceProposer.propose from the track tip (a single GEPA reflective generation, small population) + one scoring runCampaign per candidate. Elitist: the tip stays in the pool so a step never regresses below its parent (mirrors runOptimization keeping the winner as baseline when no candidate beats it). Best-scoring candidate becomes the new DAG node.
• merge = the same proposer driven with ctx.paretoParents set to the 2+ parent surfaces, which fires gepaProposer's GEPA combine-complementary-lessons crossover, then one scoring campaign on the merged surface.

The default governor is heuristicGovernor(); seed scores come from an initial scoring pass of each seed surface.

Integration choice

Single gepa-propose + one scoring campaign per step — not a full runImprovementLoop per DAG step. A full improvement loop (baseline campaign + optimization + two holdout campaigns + gate + optional PR) per single node would be enormously heavy; the DAG Governor is what controls breadth across steps (extend / branch / merge / prune), so the inner step is intentionally one small generation.

Injectable seams (unit-testable without a live model)

Both machinery halves are overridable:

• proposer — defaults to gepaProposer (needs llm + model); tests inject a pure stub.
• scoreSurface — defaults to a runCampaign pass over holdoutScenarios ?? scenarios (the "agent" = dispatchWithSurface, judges = judges); tests inject a deterministic function, or supply a stub dispatchWithSurface + judges and exercise the real runCampaign path.

Tests

src/campaign/presets/run-lineage-loop.test.ts (5 tests, all green) drives the loop with a deterministic stub agent + judge (surfaces encode their own fitness; no LLM, no sandbox) and asserts: it seeds N visioned tracks, produces a multi-node merge DAG, best() is the max score, re-running is bit-identical, the merge seam drives the crossover proposer, and required-seam validation throws. The default scoreSurface runs the real runCampaign path via an in-memory storage stub.

Observed live DAG: 3 visioned roots → a 3-parent merge node → the solve track climbs 0.11→0.31→0.51→0.71→0.9 via the real propose+campaign step seam, then plateaus and stops.

Additive: no changes to lineage.ts, run-optimization.ts, or gepa.ts. tsc --noEmit clean, biome clean, 29/29 lineage+loop tests pass.

[tangletools]

✅ Auto-approved drewstone PR — 8cc567da

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-07T17:33:42Z}