FAC-135.6 feat: sentiment vLLM-primary dispatch + admin config endpoint

[y4nder]

Summary

Expose the self-hosted vLLM classifier as the primary sentiment dispatch target with OpenAI as a per-item fallback, runtime-rotatable via a SuperAdmin admin endpoint. The vLLM deployment runs on Thunder Compute with a URL that cycles every few days as cloud credits refresh; operators now rotate the URL and model through SystemConfig without redeploying the worker.

Context

The current sentiment path calls OpenAI for every chunk at per-token cost. A self-hosted vLLM classifier (fine-tuned Gemma) is available but its URL rotates with Thunder Compute's credit cycle. The worker needs a runtime-configurable URL and the API needs a guarded admin surface to change it. This ticket covers the API-side contract, config persistence, and admin endpoint; the worker-side strategy and fallback logic live in sentiment.worker.temp.faculytics.

Scope

API (api.faculytics)

• batchAnalysisJobSchema + sentiment-worker.dto.ts extended with optional vllmConfig (https-only URL) and optional servedBy on each result item
• SENTIMENT_VLLM_CONFIG SystemConfig row seeded (enabled=false default)
• New SentimentConfigService with single-read update returning { previous, next } for TOCTOU-free audit capture
• New AdminSentimentConfigController at GET/PUT /admin/sentiment/vllm-config
  • @UseJwtGuard(SUPER_ADMIN)
  • https-only URL, cross-field validation (rejects enabled=true + empty URL/model)
  • Production rollout gate via ALLOW_SENTIMENT_VLLM_ENABLED_IN_PROD
  • Manual AuditService.Emit with before/after payload — no @Audited decorator (avoids double-audit through the interceptor)
• New audit action ADMIN_SENTIMENT_VLLM_CONFIG_UPDATE
• SystemConfig registered in AnalysisModule.forFeature; SentimentConfigService wired into PipelineOrchestratorService
• Orchestrator snapshots config once per dispatch and attaches vllmConfig to every chunk envelope only when enabled=true + URL non-empty

Tests

• Cross-boundary contract test covering the sentiment-worker envelope with vllmConfig present/absent
• Processor spec coverage for servedBy round-trip through rawResult JSONB
• SentimentConfigService spec covering read/update, { previous, next } shape, and production rollout gate
• AdminSentimentConfigController spec covering guard wiring, validation rejections, and audit emission
• Pipeline orchestrator chunking spec extended to assert vllmConfig attachment behavior

Out of scope

• Worker-side vLLM strategy + OpenAI per-item fallback (separate repo / PR)
• Admin SPA card for editing the config (separate admin.faculytics PR)
• Logprob-derived soft probabilities — empirical probe showed top-1 P≈1.0 across test samples; logprobs are preserved raw in rawResult for future calibration only
• errorClass taxonomy on servedBy failures