Broaden synthetic Jev with interpretive reading tasks

[sileod]

What changed

• Add 14 low-weight general reading skills to the existing 36-domain synthetic sampler, with no benchmark labels or special generation path. Arithmetic, chronology, entity lookup, and literal retrieval remain in the procedural segment.
• Keep general questions near 10% of the deterministic 4,000-spec draw: 606 of 5,890 questions, with every general skill present in at least 19 domains.
• Strengthen the DeepSeek critic with a per-question exact evidence quote and checks for skill match, support, and one best answer. Rename prompt files to stable names without v2.
• Correct metadata-leak validation so ordinary words such as “evidence” and “difficulty” are accepted.

Live check

A natural 120-state Albert run produced 117 validated states, 94 critic-approved states, and 119 Jev decisions in 582 seconds with one Albert key. The independent auditor answered 118/119 questions, disagreed on 12 overall and one of the 16 retained general questions, and reported no confident disagreements. The short-run rate extrapolates to about 17,700 retained decisions/day; production rate limits and sustained quota may change that.

Manual reading found skill drift and unsupported inferences that Jev/auditor agreement did not catch. A focused 16-state probe drove tighter skill instructions and the critic evidence checks. The stricter critic rejected an ambiguous negotiation implication, but missed a prediction about tomorrow based only on today's idle crew. The pilot note documents the remaining quality risk. This PR broadens the candidate distribution and improves screening; it does not certify an unattended 4,000-state export as clean training data.

Verification

• PYTHONPATH=.:src pytest -q: 217 passed
• git diff --check: clean
• Live pilot and focused probe artifacts under ignored .synthetic_runs/