Add resume-based retry mechanism to run_claude in archon-loop.sh

[surenny]

Problem

archon-loop.sh's run_claude() has no retry mechanism. If claude -p crashes mid-session (API error, rate limit, network timeout), the call silently fails via || true and the iteration continues with missing output. This wastes the work the agent already completed before the crash.

Proposed Solution

Use Claude Code's --resume <session-id> + -p to resume crashed sessions instead of restarting from scratch.

Key design:

1. Pre-assign session ID via uuidgen + --session-id on first attempt
2. Detect crash by checking if session_end event exists in the JSONL log
3. Resume with --resume <session-id> + -p "Continue your task from where you left off" — Claude picks up from the interruption point with full conversation context
4. Exponential backoff between retries (default: 30s → 60s → 120s)
5. Cap retries at 2 (3 total attempts)

Verified:

• --resume <session-id> works with -p mode (tested: session context is preserved)
• session_id is available in stream-json output's result event
• Resume prompt + existing context is enough for Claude to continue without restarting

New CLI flags:

--max-retries N        Max retries per claude -p call (default: 2)
--retry-backoff N      Initial backoff seconds, doubles each retry (default: 30)

What this does NOT retry:

• Normal completion with poor results (agent gave up, wrong approach) — not a crash
• --resume itself failing (session expired) — exhausts retry budget and returns error

Impact

All phases benefit (plan, prover, review) since retry is inside run_claude(). Parallel provers retry independently in their subshells.