[shipitdev]

Changes

• Configure TokenHarbor DeepSeek for generation and judging, with a shared locked budget ledger and cache-aware token valuation.
• Fix misleading Groq error labels, DeepSeek truncation settings and lost usage on invalid/truncated responses.
• Add supplied-evidence generator and actual API workflow collectors using the same RAGAS scoring entry point.
• Import recorded duration, usage and failure status in the initial LangSmith run records. Add metric state, eligibility and coverage feedback.
• Publish the ten-question audit, matching historical rescoring, README instructions and updated evaluation SVG.

Results and limits

• Ten new training questions produced seven answers, two withheld responses and one clarification, with zero generation errors.
• All six metric types scored for every eligible case. Component and workflow samples contain two cases each.
• Historical reference correctness still falls on four shared answers under the same DeepSeek judge. Review identifies version/scope issues, false verifier rejections and reference/judge limitations.
• Shared token value stayed at $0.295 of the $0.50 guard, including failed preflights and one unresolved reservation. The selected free route bills zero.
• These small training diagnostics use DeepSeek for both today's writer and judge. Human alignment and held-out release evaluation remain pending.

Validation

• 168 Python tests plus three subtests passed, with two optional skips.
• Separate RAGAS environment checks passed 20 tests.
• Verified native dashboard duration, tokens and error counts remotely.
• Rendered and inspected the SVG with no text overflow.
• Verified coverage, budget and redaction assertions. Original knowledge.md hash preserved.