feat(server): default exposeErrorDetails to true outside NODE_ENV=production

[bokelley]

Handler throws now emit the actual cause + handler name in dev/test/CI, so matrix runs and local dev stop losing hours on opaque 'encountered an internal error'. Production behavior unchanged — set NODE_ENV=production (or pass exposeErrorDetails: false) to get the old redacted behavior.

What changes

createAdcpServer({ exposeErrorDetails }) default:

• Before: false (always redact).
• After: process.env.NODE_ENV !== 'production' (redact only in prod).

logger.error('Handler failed', ...) now includes err.stack and handler key.

When exposeErrorDetails is on, the SERVICE_UNAVAILABLE message text is also upgraded from Tool X encountered an internal error to Tool X handler threw: <cause message> so the failure is readable at the call site — not just in server logs.

Why now

Matrix v6/v7 had 20–26 SERVICE_UNAVAILABLE rows per run with zero diagnostic. Each one was a handler throw, opaque. Validation default-on (#732) can't help — it only fires when the handler returns successfully. The fix has to be at the catch site.

Before (v7)

SERVICE_UNAVAILABLE: Tool acquire_rights encountered an internal error

After

SERVICE_UNAVAILABLE: Tool acquire_rights handler threw: Cannot find module '@adcp/client/foo'
details: { reason: "Cannot find module ...", handler: "acquireRights" }

Test plan

• Existing test returns SERVICE_UNAVAILABLE without leaking internals updated to pass exposeErrorDetails: false explicitly (documents the production-prod path).
• Full suite: 4968 pass / 0 fail.
• Typecheck clean.
• Matrix v8 after merge — expect opaque SU lines to gain specific causes, unlocking the next round of skill fixes.

🤖 Generated with Claude Code