feat: LLM retry with exponential backoff + per-call timeout

[askalf]

Summary

A 3-round deep query fires 30+ LLM calls. At a 1% per-call failure rate the overall success rate is 74%. Retries with exponential backoff bring that above 99%.

Behavior

callLLM now wraps every request in retry-with-backoff (default 3 attempts, 500ms base, 8s cap, ±25% jitter) and enforces a per-call timeout (default 120s).

Precise retry policy:

Class	Retry?
HTTP 5xx	✅
HTTP 429	✅
HTTP 4xx (non-429)	❌ — fail fast, malformed requests stay malformed
Fetch-level (network/DNS/TLS/timeout)	✅
User-initiated abort	❌ — short-circuits retry loop immediately

New API surface

• LLMError class with retriable: boolean getter
• DEFAULT_LLM_TIMEOUT_MS, DEFAULT_LLM_ATTEMPTS constants
• retry(fn, opts) in src/retry.ts — injectable sleep/random for tests, shouldRetry predicate, onRetry hook, abort-signal aware

New CLI flags

• --llm-timeout-ms=<ms> / DEEPDIVE_LLM_TIMEOUT_MS (default 120000)
• --llm-attempts=<n> / DEEPDIVE_LLM_ATTEMPTS (default 3)

LLMConfig gained optional timeoutMs + maxAttempts; resolveConfig sets defaults so library callers going through it pick them up for free (non-breaking).

Test plan

• npm run build — clean under strict: true
• npm test — 159 pass (up from 141 pre-branch), 0 fail
• 21 new assertions: retry helper math/behavior (12) + LLM integration covering first-try-OK, 500→500→200, 429→200, 400 does-not-retry, 401 does-not-retry, exhausted 5xx throws LLMError, per-call timeout on hung server, user-abort short-circuits (9)