feat(heartbeat): automatic diagnostic retry for failed runs

[Noesis-Boss]

Summary

When a run fails due to an adapter error (network timeout, rate limit, API error, etc.), Paperclip now automatically classifies the failure and can retry it up to 2 times using the heartbeat service as a diagnostic agent.

What changed

New column: diagnosticRetryCount

• Added to heartbeat_runs via migration 0046_curious_deadpool.sql
• Tracks how many diagnostic retries have been attempted for a run
• Max 2 retries (count=0 first attempt, count=1 second attempt, count=2 stops)

New function: maybeEnqueueDiagnosticRetry in heartbeat.ts

After every failed run in finalizeCompletedRun, this function:

1. Checks if diagnosticRetryCount < 2
2. Classifies the failure by parsing error, exitCode, and signal
3. If canRetry = true (transient errors only), enqueues a diagnostic retry run
4. Updates diagnosticRetryCount on the new run

Failure classification (classifyFailure)

Each failure is categorized and marked retryable or not:

Category	Retryable	Reason
process_killed	✅	SIGKILL/SIGTERM
rate_limit	✅	API rate limit
timeout	✅	Network timeout
network_error	✅	Connection reset, EOF errors
eof_error	✅	Empty response
not_found	❌	Missing file
permission_denied	❌	Permissions issue
oom	❌	Out of memory
unknown	❌	Unclassified

No changes for process_lost

The existing processLossRetryCount retry mechanism is unchanged.

Testing

• Migration applied, server restarts successfully
• TypeScript compiles (pre-existing upstream errors unrelated to this PR)