fix(attention): keep public metrics finite and numeric

[rudycelekli]

Problem

Closes #782.

Attention feeds can admit JSON numeric overflow (1e400 becomes infinity), booleans coerced to counts, and legacy nonfinite metrics. An oversized integer can drop an entire otherwise usable feed. These values can leak into public metrics from primary, supporting, or restored observations.

Change

Use one metric normalizer across those three paths. Retain finite nonnegative numbers, including measured zero; exclude booleans, nonfinite values, and values that overflow conversion. Preserve feed items and other valid metrics.

Verification

The regression uses the native attention collector and restoration callers: two failing tests before the fix, six passing focused tests after it. Strict JSON serialization is asserted.

All six required commands pass at this exact signed head in a clean detached checkout with its own editable dev environment and initialized recursive submodule:

• ruff check .
• ruff format --check .
• benchmark-radar normalize-catalog
• benchmark-radar classify
• benchmark-radar build-data-release
• pytest -q — 1,642 passed

The regenerated catalog contains 3,212 records from five sources, with matching index/shard/archive identities. Python 3.12.11; bundled Node runtime used for native site checks. Hosted CI status will be reported separately.

Contribution

Prepared with AI assistance (gpt-6.1-sol), reviewed against native callers and regression evidence. The SSH-signed commit includes a DCO signoff. No source data, generated data, or paper changes.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[rudycelekli]

330226 gpt-6.1-sol

[ktwu01]

Closing as not planned. One author now has a cap of 10 open issues + PRs, and this account had 56 open PRs and 11 open issues, mostly paired with each other. We merged the three changes that fix silent data loss in the corpus (#723, #774, and #773 once rebased) and are closing the rest so review stays possible.

This is not a verdict that the change is wrong. If you think this one matters, reopen it later when you are under the cap, one at a time, with a reproduction from a real user or caller rather than a hypothetical edge case.

不是否定这个改动本身。为保证评审能跟上，同一作者最多同时保留 10 个未关闭的 issue/PR；该账号曾有 56 个未关闭 PR 和 11 个 issue。已合并修复语料静默丢失的三个改动，其余先关闭。若认为本条重要，请在低于上限后逐个重新打开，并附真实用户或调用方的复现。

---

Generated by Claude Code