feat(improvement): source both proposers uniformly from agent-eval/campaign

[drewstone]

What

agent-eval 0.97.0 finished the Proposer migration: the optimize-a-surface proposers are now uniform factories on /campaign — gepaProposer (prompt surface) and skillOptProposer (skill surface), both xProposer(opts): SurfaceProposer. This points the three consumers at that one source instead of the split they carried during the migration.

• src/improvement/improve.ts — the documented surface → default proposer map now imports both gepaProposer and skillOptProposer from /campaign (was: gepa from /contract, skillOpt from /campaign). selfImprove + the SelfImprove* loop types stay on /contract (the loop contract, not a proposer). This closes the call-time-crash risk the migration left: a drifted proposer export now fails the build.
• src/lifecycle/prompt-generator.ts — drop the /contract special-case; gepaProposer comes from /campaign alongside the ProposeContext/GenerationRecord types already imported there.
• src/lifecycle/skill-generator.ts — refresh the production-wiring docs to name skillOptProposer (the uniform factory) instead of the lower-level runSkillOpt internal. The distill + diverse-seed arms are untouched — that is the generator value-add the proposer cannot do.
• src/improvement/improve.test.ts (new) — drives the real improve() to a gate decision for both default surfaces (prompt → gepaProposer, skills → skillOptProposer) so a drifted proposer export fails the build, not a caller. Deterministic + offline: gate: 'none' forces generations = 0 (no reflection LLM call); the stub agent reports token-bearing cost so the backend-integrity guard sees a real backend. Also asserts the no-default config surfaces (tools/mcp/hooks/code) still fail loud with ConfigError.

Dependency

@tangle-network/agent-eval peer + dev pin → >=0.97.0 <1.0.0; agent-runtime → 0.75.0; docs/api regenerated + version pin updated.

Verify (all green, run against real 0.97.0)

• pnpm run build — clean (DTS emitted)
• pnpm run typecheck (src + examples) — clean
• pnpm test — 115 files, 1113 passed, 1 skipped, 0 errors. The feat(lifecycle): profile-artifact lifecycle as a runtime primitive — generate/eval/promote/maintain skills, tools, MCPs, hooks, subagents for free #267 closed-loop proof (src/lifecycle/closed-loop.test.ts) and the new improve.test.ts both pass.
• pnpm run lint (Biome) — 0 errors
• pnpm run docs:check — docs/api regenerated (idempotent) + freshness gate OK
• Runtime probe against installed agent-eval@0.97.0: gepaProposer/skillOptProposer resolve from /campaign to real factories (.kind = gepa/skill-opt, .propose present; skillOptProposer also exposes .proposePatches); selfImprove resolves from /contract.

[tangletools]

✅ Auto-approved PR — c0bba96b

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-22T22:49:46Z}