refactor(bench): rename gepa-refine → improve-prompt (name by purpose, not method)

[drewstone]

gepa-refine was a bad name twice over: named after the method (GEPA, which is swappable) instead of the purpose, and "refine" collides with the inner within-run loop — but this file is the outer cross-run improvement loop.

Naming convention this establishes (so the whole improvement system reads clearly):

• improve- / optimize- = the OUTER loop — across runs, optimize a surface, held-out gated.
• refine- = the INNER loop — within a run, k rounds over one persistent artifact.
• name by surface/lever, not by algorithm: future siblings are improve-skill.ts, improve-tool.ts, improve-config.ts.

This file is the OUTER loop over a string surface (MutableSurface = string | CodeSurface) — here the refine directive. The same loop optimizes skills / inter-agent messages (also strings) and code (CodeSurface) by swapping the surface, not the loop. GEPA stays in the docstring as the string-surface optimizer it uses.

Mechanical + safe: it's a CLI entrypoint, nothing imports it, so no code references change. git mv + the npm script (gepa-refine → improve-prompt) + help text + HARNESS.md + 2 doc refs. bench/package.json re-validated as JSON; entrypoint intact.

🤖 Generated with Claude Code

[tangletools]

✅ No Blockers — eeddd6df

Readiness 89/100 · Confidence 85/100 · 1 finding (1 low)

	deepseek	glm	aggregate
Readiness	89	89	89
Confidence	85	85	85
Correctness	89	89	89
Security	89	89	89
Testing	89	89	89
Architecture	89	89	89

Full multi-shot audit completed 5/5 planned shots over 7 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 5/5 planned shots over 7 changed files. Global verifier still owns final merge decision.

🟡 LOW Stale line-number reference in docs/roadmap-rsi.md:46 — docs/roadmap-rsi.md

Line 46 references improve-prompt.ts:273-281 for the heldoutSignificance(pairHoldout(...)) ship-gate pattern. The actual call site is at lines 589-602 (the block comment starts at :589, the pairHoldout call at :601, heldoutSignificance at :602). Lines 273-281 contain unrelated AppWorld driver JSON parsing code. The reference was already stale in the original gepa-refine.ts filename and was carried forward unchan

---

_{tangletools · 2026-06-05T23:22:08Z · trace}

[tangletools]

✅ Approved — 1 non-blocking finding — eeddd6df

Full multi-shot audit completed 5/5 planned shots over 7 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 5/5 planned shots over 7 changed files. Global verifier still owns final merge decision.

Full immutable report for this review: trace

Summary comment for this run: full summary

---

_{tangletools · 2026-06-05T23:22:08Z · immutable trace}