hotfix: run_claude diagnostics + empty-diff tracking (Claude CLI path)

[telleroutlook]

Why

PR #260 added Codex-path diagnostics to run_codex_exec in worker.py. Production log 2026-07-02 13:42 shows the actual empty-diff cause is different:

INFO no diff (nothing changed)
INFO empty-diff counter for telleroutlook/claude-bot-go#8: 11
WARNING Claude exited with returncode=1
WARNING Job 264 blocked (empty_diff_detected): Empty diff after 11 attempt(s)

Claude CLI (not Codex) is the one returning exit=1 with no diff. That path lives in claude_ops.run_claude, which PR #260 didn't touch.

What this changes

Only run_claude in claude_ops.py. No control-flow changes. Pure observability.

• Timing: runtime_secs per invocation
• Result summary: cwd, model (Claude alias or Codex model), returncode, stdout_bytes, stderr_bytes
• Suspicious-short-output warning (returncode==0 && stdout<50 chars)
• Worktree diff probe: git status --porcelain + git diff --numstat HEAD
• Per-worktree empty-diff streak counter (module-level dict), reset on real change
• At streak >= 2, emit a claude-bot-go#23 pointer so the Go-side native implementation has a hook

What this does NOT do

• Does not change return values or callers
• Does not add auto-rebuild logic (that is claude-bot-go#23 in Go bot)
• Does not touch token parsing, error handling, or timeout paths
• Does not change any config

Verification

• python3 -m py_compile claude_ops.py OK
• Diff is +54 -0 additive, no behavioral changes

Follow-up (autonomous)

• Downstream: claude-bot-go#22 (empty_diff auto-unblock) + feat: patrol redesign — fix queue-blocking + add cross-repo org coherence patrol with opus #23 (native diff verify) + fix: feat: patrol redesign — fix queue-blocking + add cross-repo org coherence patrol with opus (#23) #24 (zero-pending watchdog)
• Once deployed, next hour's log will show which provider (Claude vs Codex) is failing, with what runtime, returncode, and output shape — real fix follows from data

Refs: claude-bot-go#23, #260