Version 3.3.3

3.3.3 is a focused infrastructure fix. There are no UI changes. The change is on the backend recovery sweep that resurrects scoring jobs lost to a process restart. It now gives up gracefully on jobs that can't be scored — instead of re-attempting them forever and quietly accumulating compute cost.

Behind the scenes

• Recovery sweep gives up after three tries. Pre-fix, the periodic recovery sweep would re-attempt scoring on any job whose match_score was still NULL past the safety threshold. If a job hit a permanent failure mode (e.g., the skill ontology produced zero matches), the sweep would re-fire the same Anthropic call every 60 minutes — forever. Now the job exits the eligibility set after three attempts and is marked as scoring-failed; the dashboard can render the failure state plainly.

• Skipping jobs that already have a score tier. The sweep also now skips jobs where the embedding pipeline already produced a preliminary score — that's already-visible scoring data, no need to re-fire the deep orchestrator on top.

Why it matters

This is the unblocker for re-enabling the recovery sweep in production (currently disabled by a kill switch). Once it's safe to flip back on, normal stuck-job recovery resumes — without the runaway loop. The user-visible effect should be: jobs that genuinely lost their scoring task get rescored within five minutes; jobs that can't be scored stop pretending they will be eventually.

---

Pure plumbing. Nothing to click. Background scoring just got more honest about what it can and can't recover.