Routines: execution fails silently when sandbox/Docker unavailable

[sergeiest]

Component: src/agent/routine_engine.rs, src/sandbox/, src/agent/scheduler.rs

When routines dispatch full_job execution but the required infrastructure is unavailable (Docker not installed, sandbox misconfigured, container runtime inaccessible), the job fails with zero feedback:

• started_at: NULL
• completed_at: NULL
• status: failed
• No error message in routine_runs, agent_jobs, or any log

This is not limited to Docker specifically -- any environment where the execution backend is unavailable (managed infra without Docker, containers without required runtimes, misconfigured sandbox policies) results in silent failure. The user must inspect the database manually to discover failures.

Impact: full_job is the only routine action type that can execute tools, but it silently fails whenever execution infrastructure is missing. On managed infrastructure without Docker, ALL routine types are broken -- lightweight can't use tools (#1), full_job silently fails (#2).

[sergeiest]

Logs / Evidence:

Philipp (Slack):

Switched to full_job -- silently fails. Changed blog_auditor to action_type: full_job. The routine dispatches a job (Dispatched job b4e685e4... for full execution with tool access), but the job record shows started_at: NULL, completed_at: NULL, status: failed. No error message anywhere -- not in routine_runs, not in agent_jobs, not in any log file.

Root cause: sandbox.enabled = true in config but Docker is not installed on the managed infrastructure. The job can't start a sandbox container. There is zero feedback about this -- the user has to discover it by inspecting the database manually.

Pierre (PDF p10):

"Creating the job seemed to work, but forcing the job to run didn't yield to anything."

Henry (PDF p11) -- Routine dispatch log followed by failure:

23:33:18.64/INFO ironclaw::agent::routine_engine Dispatched full job for routine
routine=ai-news-telegram-5min job_id=3e0d3580-42e7-4fe8-aeeb-bccbb3934c35
max_iterations=10

Jobs list shows 3 failed runs: 3e0d3580, 50e53664, ea255b69 -- all status failed.

[sergeiest]

Suggestion (from Henry): When execution fails, surface clear error messages. Detect infrastructure availability at startup and warn. Consider fallback to in-process execution when the sandbox is unavailable.