fix: parse Claude.ai privacy export with messages key and sender field (#677)

[mvalentsev]

Closes #677.

Claude.ai's Settings > Privacy > Export Data produces a conversations.json that _try_claude_ai_json can fail to parse for two reasons:

1. Key variant -- the privacy-export guard only checks for "chat_messages" in each conversation object. If an export uses "messages" instead (as described in fix: handle large claude.ai exports and multi-conversation "messages" key #676), the guard misses it. The array of conversation objects falls through to the flat-messages parser, which expects {role, content} dicts and silently skips the {uuid, name, messages} objects. The function returns None and the raw JSON gets filed as a single plain-text drawer.

2. Author field variant -- some privacy exports may use "sender": "human"/"assistant" instead of "role" (as described in Support claude.ai privacy export format with sender field #605). When the outer guard matches but the inner loop only checks item.get("role"), every message is skipped, producing an empty transcript.

Both failure modes result in the behavior reported in #677: a multi-MB file mined as one drawer classified "emotional."

Changes (2 files, +129/-21):

mempalace/normalize.py:

• Outer guard now accepts both "chat_messages" and "messages" keys
• Inner extraction: convo.get("chat_messages") or convo.get("messages", [])
• Author field: item.get("role") or item.get("sender", "") handles both variants
• Text fallback: tries item.get("text") when content blocks are empty
• Per-conversation transcripts instead of concatenating everything into one blob
• Shared _collect_claude_messages() helper to deduplicate extraction logic

tests/test_normalize.py:

• 6 new tests: messages key, sender field, text fallback, per-conversation separation, empty-conversation skipping

Scope note: this PR only touches the normalizer. It does not change convo_miner.py -- the 10 MB MAX_FILE_SIZE limit is unrelated to the parsing failure in #677 (the reported file is 8 MB) but is addressed separately in #605 and #676.

Related PRs: #605 (carlito1979) adds sender support, text fallback, and raises MAX_FILE_SIZE. #676 (z3tz3r0) adds messages key detection, per-conversation separation, and also raises MAX_FILE_SIZE. Both touch convo_miner.py which this PR does not. This PR combines the normalize.py parsing fixes from both approaches and adds the missing cross-coverage (neither PR alone handles both the key variant and the author field variant).