feat: agent self-correction via validation feedback loop

[nicknisi]

Summary

• Add fast typecheck/build validation that runs between agent turns, giving the agent structured feedback to self-correct within the same session (up to 2 retries)
• Auto-detect build systems across ecosystems: JS (package.json), Go (go.mod), Elixir (mix.exs), .NET (*.csproj), Kotlin/Java (build.gradle) — interpreted languages pass through silently
• Unify eval executor with production runAgent so evals exercise the actual retry path
• Track three-tier pass rates: first-attempt, with-correction, with-retry

Why

The installer ran its agent as a single-shot operation — when validation caught fixable issues, the results went to the user, not back to the agent. The agent never got a chance to fix its own mistakes.

Eval results (14 frameworks, --state=example):

Metric	Value
First-attempt pass rate	92.9%
With-correction pass rate	100%
Self-corrected scenarios	1 of 14
Quality score	4.5/5

Architecture

Agent writes code
    ↓
Typecheck (tsc --noEmit, TS only, ~5s)
    ↓
Pass? → Build (auto-detected per ecosystem)
    ↓         ↓
    ↓    Pass? → Full validation (env vars, files, patterns)
    ↓      ↓
    ↓   Format errors → yield back into same SDK conversation
    ↓      ↓
    ↓   Agent fixes (retains full context, max 2 retries)

The retry loop uses an async generator that yields follow-up user messages into the SDK's query(). The agent retains full conversation context.

Changes

Quick checks (src/lib/validation/quick-checks.ts): Typecheck + build as composable steps. Short-circuits on typecheck failure. quickCheckValidateAndFormat shared between production and evals.

Multi-ecosystem build detection (src/lib/validation/build-validator.ts): detectBuildCommand checks package.json, go.mod, mix.exs, *.csproj, build.gradle. Returns null for interpreted languages.

Retry loop (src/lib/agent-interface.ts): Async generator yields correction prompts on validation failure. Promise-based turn coordination. Exports AgentRunConfig + onMessage hook for evals.

Evals (tests/evals/agent-executor.ts): Delegates to production runAgent. Three-tier success criteria: first-attempt (80%), with-correction (90%), with-retry (95%). --no-correction flag.

Validator composability (src/lib/validation/validator.ts): Exported validatePackages, validateEnvVars, validateFiles, validateFrameworkSpecific with return-based signatures.

Notes

• dotnet eval scenario disabled (broken SDK)
• Quality grader JSON parsing fixed (greedy regex matched braces inside <thinking> tags)