fix: detect service unavailability and fail fast with clear error

[nicknisi]

Summary

Fixes the installer hanging for ~9 minutes with no useful feedback when the Claude API is down (reported by MG in Slack).

Root cause: The SDK returns subtype: 'success' with is_error: true when API retries are exhausted. Our code only checked subtype, so it treated the error as success and proceeded with validation retries — 3 cycles of 10 retries each (30 total API calls, ~9 minutes) before surfacing a raw JSON error.

Before: 30 API calls over ~9 min, user sees: Claude Code returned an error result: API Error: 500 {"error":{"type":"internal_error","message":"An unexpected error occurred"}}

After: 10 API calls over ~3 min (first cycle only), user sees: The AI service is temporarily unavailable. Please try again in a few minutes.

Also adds clear error messages for other failure modes that previously showed raw/cryptic errors:

Error	Before	After
API 500/503	Raw JSON after 9 min	"AI service temporarily unavailable" (~3 min)
429 rate limit	Raw JSON	"AI service is currently rate-limited"
Network failure	ECONNREFUSED / ETIMEDOUT	"Could not connect to the AI service"
Process crash	process exited with code 1	"AI agent process exited unexpectedly" + --debug hint

Changes

• agent-interface.ts: Check is_error on result messages, classify 500/server_error/429 as SERVICE_UNAVAILABLE, add abortRetries flag to skip validation retries on fatal errors
• cli-adapter.ts: Detect service, rate limit, network, and process exit errors; show clean messages instead of raw API JSON
• headless-adapter.ts: Emit structured error codes (service_unavailable, rate_limited, network_error, process_error) for JSON consumers
• agent-interface.spec.ts: 4 new tests covering is_error detection, service classification, and retry abort

Test plan

• pnpm test passes (1538 tests)
• pnpm typecheck passes
• pnpm build passes
• New tests cover: is_error detection, SERVICE_UNAVAILABLE classification, non-service error fallback, validation retry skip