Add --retry flag to re-run only failed or timed-out agents

[that-github-user]

Problem

When some agents fail or time out, the user must re-run the entire ensemble to get more results. This wastes time and API credits by re-running agents that already succeeded.

Common scenarios:

• 1 of 5 agents timed out; user wants a 5th successful result
• 2 agents errored due to a transient issue (rate limit, network blip); user wants to retry just those
• User wants to top up: had 3 agents, wants 2 more without losing the first 3

Proposed Solution

thinktank run --retry

Re-run only the agents that failed, timed out, or errored in the most recent run:

$ thinktank run --retry
Found last run: 5 agents (2 timed out, 1 error)
Re-running 3 failed agents...

  Agent 4  (was: timeout)  →  ✓ done  (tests passed)
  Agent 5  (was: timeout)  →  ✓ done  (tests passed)  
  Agent 6  (was: error)    →  ✓ done  (tests failed)

Merging with 2 successful results from previous run...

The merged result set (original successes + new attempts) goes through convergence analysis together.

thinktank run --top-up <N>

Add N more agents to the existing result set and re-run convergence on the combined results:

$ thinktank run --top-up 2
Adding 2 more agents to last run (was: 3 agents)...

Acceptance Criteria

• --retry identifies failed/timeout/error agents from the latest run file
• Re-runs only those agents in new worktrees (successful agents are not re-run)
• Results are merged with the surviving successful agents from the prior run
• Merged result set goes through convergence analysis as a single combined set
• --top-up <N> runs N additional new agents and merges with prior run
• If no prior run exists, --retry and --top-up print a clear error
• --run <N> can be combined to retry from a specific historical run
• Merged results saved as a new timestamped run file (prior run is not modified)

[that-github-user]

Fixed in #115. --retry re-runs only failed/timed-out agents. Generated by thinktank Opus.