fix(uploads): SkillForge text fallback + queue self-heal

[criptogus]

Summary

Diagnostic round 2 surfaced two production failures behind the freshly-shipped REST upload path. Same root areas, two different layers.

NEW-1: SkillForge author 0/1 on every inline upload

• processBulkUpload was returning uploaded:0, failed:1 for every call. Both structured-output fallback models died:
  • openai/gpt-4o-mini → Bad Request (provider strict mode rejecting the JSON schema generated from PackageDraftSchema's open record<string, any> slots).
  • google/gemini-2.5-flash → No object generated: response did not match schema.
• The catch swallowed the upstream body — only e.message reached logs, so the symptom was a generic "Bad Request" with no signal.

Fixes

• describeAttemptError extracts e.text (NoObjectGeneratedError), e.responseBody (HTTP errors), and e.cause.message. Real provider body now lands in logs.
• After all structured attempts exhaust, a text fallback runs: plain generateText asking for raw JSON, then parsed locally and validated with PackageDraftSchema. Tolerant extractor handles bare JSON, ```json fences, and brace-balanced first-object recovery. Loses provider-side schema enforcement but yields a draft instead of a fatal error.

NEW-2: queue zombies — uploads stuck in queued indefinitely

• Worker hardcoded attempts: 1 on every claim, so MAX_ATTEMPTS gating could never trigger.
• A job that crashed mid-flight (Vercel 60s budget, network flap) stayed in processing forever with no recovery path.

Fixes

• Each drain pass scans processing jobs with started_at < now-2min:
  • attempts < 3 → flip back to queued, clear started_at.
  • attempts >= 3 → mark failed with an "abandoned" reason so the user sees a real outcome.
• Claim increments attempts from the prior value instead of overwriting.
• Return value adds requeued so cron logs surface the healing.

Out of scope (still TODO)

• BUG-3 MCP-side Bearer sas_… rejection (suspected separate auth context inside the MCP tool, not the route). Will tackle next.
• BUG-1 phantom super-agent CLI on npm — needs a publish or a snippet revision; not a code-in-this-repo fix.

Test plan

• POST /api/packages/upload with a real file returns uploaded:1, failed:0 — text fallback should catch any remaining structured-output strictness issues.
• When all models fail, response results[0].error now contains text=… or body=… snippets pointing at the upstream cause, not just "Bad Request".
• Manually UPDATE package_upload_jobs SET status='processing', started_at=now()-interval '5 minutes' WHERE …; hit /api/jobs/drain-upload-queue?secret=…. Job flips back to queued (attempt 1) or failed with "abandoned" (attempt 3).
• Drain response JSON now includes requeued counter.
• Existing 14h-old zombie jobs reachable from /account/packages clear within one cron tick after this deploys.

https://claude.ai/code/session_019gMoupKKTVydpNwiiACQRd

---

Generated by Claude Code