generator: correct failure feedback and paper semantics

[huangyz0918]

Summary

Fix failure-discovery behavior found while checking Sections 2.4 and 3.1 and Algorithm 2 of the paper: https://arxiv.org/html/2604.23099v1

• Count failure feedback, not successful answers, as topic reward; connect the normal sequential generate() → update(score) loop to topic statistics.
• Explore each untried topic before applying the UCB formula.
• Select TSS anchors independently of the chosen topic.
• Describe high error probabilities as hard examples and failure rates in generation prompts, rather than reversing them into success rates.
• Honor the caller's hard-selection threshold and rank using the latest posterior's high failure scores.
• Detach empty-observation encoder kernels before NumPy conversion so trainable Matérn/RBF parameters do not cause a runtime error.

Known limitation

The current score-only update(score) API cannot condition the GP on a newly generated input: it lacks that input's embedding or kernel covariances. This PR corrects the misleading docstrings and topic feedback but does not invent a generated-point covariance or assign its score to an anchor. Full generated-input GP conditioning remains a separate API gap; this PR does not claim complete Algorithm 2 conformance.

Verification

• Existing repository suite: 123 passed, including PyTorch paths.
• Local probes cover sequential topic feedback, cross-topic anchors, UCB exploration, prompt score direction, threshold forwarding, posterior-based hard selection, and empty/nonempty linear/Matérn/RBF posteriors.
• Only production bug fixes and their inline documentation corrections are included. Additional validation code stays local; no credentials or live API calls are involved.