feat(0.11.0): defineAgent + surfaces-driven adapters + outcome measurement

[drewstone]

Summary

The substrate redesign. Closes every universal failure flagged by the audit of the four per-agent draft PRs (now closed):

1. Subject-routing dead code → typed `FindingSubject` (agent-eval 0.30.0) + `resolveSubjectPath(subject, surfaces, repoRoot)` resolves to a real file path. No `startsWith(...)` prose matching.

2. ImprovementAdapter has no apply → `createSurfaceImprovementAdapter` ships first-class apply with three modes:

   • `write` — `git apply -p0` in-place; operator reviews via `git diff`
   • `open-pr` — branch + commit + push + `gh pr create`
   • `none` — report-only
     Race-checked via SHA-256 of the file content the patch was drafted against.

3. No outcome measurement → `measureOutcome(result, opts)` re-runs the cohort after apply, computes composite delta, optionally rolls back applied paths on regression.

4. Fabricated file paths → `validateSurfaces` runs at `defineAgent` time and throws `AgentManifestError` listing every missing surface. Broken manifests can't ship.

5. Per-vertical ImprovementAdapter duplication → `createSurfaceImprovementAdapter(opts)` is the ONE adapter every vertical uses. Customization moves to the manifest's `surfaces` + `autoApply`, not new adapter code.

6. Knowledge proposals lacking lint → `createSurfaceKnowledgeAdapter` wraps `applyKnowledgeWriteBlocks` with an optional `lintAfterApply` hook.

API surface

New sub-export `@tangle-network/agent-runtime/agent` ships:

• `defineAgent<TPersona, TRunOutput>(manifest)` — typed, validated factory
• `AgentSurfaces`, `validateSurfaces`, `resolveSubjectPath` — surface map + Subject→Path resolver
• `createSurfaceImprovementAdapter(opts)` — LLM-drafted patches + `git apply` / `gh pr create` apply
• `createSurfaceKnowledgeAdapter(opts, deps)` — agent-knowledge integration with post-apply lint
• `measureOutcome(result, opts)` — before/after cohort delta + rollback
• `AgentManifestError` — fail-loud manifest validation

Why the redesign

The audit of the four draft PRs found the same structural failures across every one: subject grammar unenforced, no apply on the improvement side, no outcome metric, fabricated paths, per-vertical glue that doesn't scale past 4 verticals — let alone the 1000s the substrate needs to serve. Fixing those in-place across 4 PRs would have shipped the same theater; fixing the substrate fixes them by construction for every future agent.

After this lands, per-vertical PRs collapse from ~700 lines of glue to a ~50-line `defineAgent({...})` call + the substrate's default adapters.

Test plan

• `pnpm test` — 124/124 pass (27 new under `tests/agent.test.ts`)
• `pnpm typecheck`
• `pnpm build` — emits new `agent.{js,d.ts}` sub-export
• CI install gated on agent-eval 0.30.0 publish — draft until then.

Gates

• agent-eval#61 (`feat(0.30.0): FindingSubject typed grammar`) merges + 0.30.0 publishes
• Then this PR's `@tangle-network/agent-eval ^0.30.0` dep resolves and CI unblocks