[shipitdev]

Changes

• Add optional, fail-soft LangSmith traces around each request, embedding, exact search, reranking, context selection and Groq calls.
• Record actual tokens, cache use, sources, verification/repair decisions and released outcomes. Keep trace links and diagnostics available for failed requests.
• Publish saved runs as LangSmith experiments with free contract checks and optional matching RAGAS scores. Keep the dataset stable when answers change.
• Add seeded training sampling, a hard HTTP-attempt cap, saved checkpoints and stop-on-quota behavior.
• Replace the spinner/checklist panel with live activity and early source excerpts. Keep verified release checks, add retained activity history, local fonts and responsive navigation.
• Publish a redacted observability result and document setup and the streaming API contract.

Measured batch

• Twelve fixed training tuning questions. Labeled article selected for 11/12.
• Three released answers. Citation-format checks passed for 3/3.
• Nine provider quota errors. Fifteen HTTP attempts and 33,230 reported tokens.
• No new semantic RAGAS scores or held-out evaluation. Quota errors are separate from answer-quality failures.

Checks

• Local Python suite passed with 151 tests, 3 subtests and one optional RAGAS integration skip.
• Eight frontend contract tests passed. Typecheck, lint and production build passed.
• Controlled browser transport verified progress before approval, source previews, OpenUI rendering and session reset at 1440, 768 and 390 pixels, without horizontal overflow or page errors.
• Real LangSmith connection, saved experiment publication and a detailed cache-hit retrieval trace were verified. Browser fixtures made no model calls.

Limits

• The quota-limited run does not establish answer correctness. New semantic judging awaits provider quota.
• Private traces include questions, evidence and drafts. Credentials and raw records remain outside Git.