fix: guard against null entries in LLM extraction response

[dingguagua996-stack]

Summary

When the smartExtraction LLM returns malformed JSON with null entries in the memories array (e.g. {memories: [null, {...}]}), accessing raw.category throws:

TypeError: Cannot read properties of null (reading 'category')

This crash kills the entire auto-capture pipeline, producing the log:

memory-lancedb-pro: capture failed: TypeError: Cannot read properties of null (reading 'category')

Root Cause

In src/smart-extractor.ts, the candidate processing loop iterates over result.memories without checking if individual entries are valid objects:

for (const raw of result.memories) {
  const category = normalizeCategory(raw.category ?? ""); // 💥 raw is null!
}

Some LLM models (especially under load or rate limiting) occasionally return arrays with null entries mixed in with valid objects.

Fix

Add a null/type guard at the top of the loop to skip invalid entries gracefully:

if (!raw || typeof raw !== 'object') {
  invalidCategoryCount++;
  continue;
}

Evidence

Observed 7 occurrences across multiple days in production:

2026-03-21T16:14:36 capture failed: TypeError: Cannot read properties of null (reading 'category')
2026-03-21T20:26:05 capture failed: TypeError: Cannot read properties of null (reading 'category')
2026-03-22T20:06:13 capture failed: TypeError: Cannot read properties of null (reading 'category')
2026-03-24T02:46:33 capture failed: TypeError: Cannot read properties of null (reading 'category')

Files Changed

File	Change
src/smart-extractor.ts	Add null/type guard for candidate entries in extraction loop

[AliceLJY]

看了代码，fix 方向没问题，smart-extractor 里确实缺 null guard。

我们刚发了 v1.1.0-beta.10，main 有比较大的变动，麻烦 rebase 到最新 master。另外补一个回归测试：在 test/smart-extractor-branches.mjs 里加一个 memories: [null, validEntry] 的 case，确认不会炸。

rebase + 补测试后可以合 👍