Compact inflated AI session transcripts

[jsgrrchg]

Summary

This PR fixes the root cause of inflated AI session transcripts in .neverwrite/sessions/*/transcript.jsonl.

NeverWrite already used session-meta.json and index.json as the logical source of truth, but the physical JSONL transcript could keep accumulating obsolete versions of updated tool/status messages. Long ACP sessions, especially Codex sessions with many tool/status updates, could therefore leave hundreds of MB of duplicated transcript rows even when the indexed conversation was much smaller.

Changes

• Adds shared transcript compaction in crates/ai/src/persistence.rs for all AI runtimes, not just Codex.
• Rewrites inflated transcript.jsonl files using only the messages currently referenced by index.json.
• Rebuilds transcript offsets, lengths, and hashes after compaction.
• Uses temporary files, backups, and compact-state.json so interrupted compactions can be rolled back safely on the next load/save.
• Compacts on save when obsolete transcript bytes are significant, now it uses 6 times less space.
• Repairs old inflated sessions on load/restore/page fetch before reading transcript pages.
• Avoids compacting summary-only history listing, so opening Chat History metadata does not do unnecessary heavy work.

Impact

The user-visible chat history is preserved. This only removes obsolete physical JSONL rows that represented intermediate versions of messages already superseded by the transcript index.

This should reduce disk growth, I/O, memory pressure, and crash/freeze risk from long AI conversations while keeping restore/history behavior unchanged.

Validation

• cargo test -p neverwrite-ai
• cargo build -p neverwrite-native-backend
• git diff --check