feat: support new Gemini CLI JSONL format and filter metadata

[bet4it]

This PR adds support for the new Gemini CLI JSONL session transcript format and filters metadata blocks to prevent FTS/title indexing pollution.

Changes

1. JSONL Format Parser: Added a JSONL parser in packages/core/src/parsers/gemini.ts that handles streaming JSON lines (supporting direct message entries and the $set operator).
2. Metadata Filtering: Implemented XML-style tag filtering for <session_context> blocks in user inputs to prevent metadata pollution.
3. Session Paths: Updated packages/core/src/sync/source-paths.ts to include .jsonl patterns in watch lists and fix nested tmp/tmp workspace folder resolution.

[graydawnc]

Thanks for this — the format support is real and needed (verified against gemini-cli source: JSONL landed upstream in 0.39.0, and <session_context> injection is exactly as you described). I reproduced the full matrix locally with real session files from 0.38.2 (legacy .json), 0.49.0 (JSONL), and a 0.49.0 --resume migration of a 0.38.2 session, and found a few gaps against upstream's own loader semantics (chatRecordingService.ts → loadConversationRecord). I've pushed a commit (76876e3) on top of your branch that addresses them:

Parser (vs upstream replay semantics)

• $rewindTo records were ignored — rewound-away messages stayed in the index forever. Now dropped (target + everything after), matching upstream.
• $set.messages is a checkpoint upstream (messagesMap.clear() + rebuild); the merge approach kept messages that a checkpoint had removed. Now clear-and-rebuild.
• Same-id snapshots (tool-call status updates) are full replacements upstream; shallow-merge could keep stale fields a later snapshot legitimately dropped. Replay is now keyed off a Map, which also removes the O(n²) findIndex per record.
• <session_context> stripping now applies to user messages only, so assistant output quoting the tag isn't mangled.

Sync layer (format-transition hazards)

• Resuming a legacy session in ≥0.39 migrates it to a sibling .jsonl and leaves the old .json in place — same sessionId in both. Indexing both made them clobber each other's row via UNIQUE(session_uuid) on every scan (reproduced: syncAll never converged, file_path flip-flopped each round). A .json with a .jsonl sibling is now skipped.
• A rewind followed by enough new turns keeps the replayed count ≥ the stored row count with the head unchanged, which the append/rewrite heuristic misread as append — rewound rows survived with colliding seq values. classifySync now also checks the stored tail uuid for gemini.
• Bumped the gemini index version to v2 and made a version change force the rewrite path (append-mode INSERT … DO NOTHING re-visits files after a bump but updates nothing), so sessions indexed before the stripping re-derive their content.

Also confirmed your tmp/tmp fix addresses a real bug — the workspace dir under ~/.gemini/tmp/ is named after the cwd's basename, so running gemini inside any directory named tmp creates ~/.gemini/tmp/tmp/, and the old root normalization then re-nested the root and made every gemini session invisible. Nice catch; I added a regression test for it.

All new tests fail on the pre-fix code and pass after; core suite is green (404 passed). One known limitation left as-is (pre-existing, applies to the legacy format too): an in-progress same-id snapshot indexed mid-turn by the live watcher won't have its content updated by later snapshots until a rewind/truncation or an explicit "Refresh from source" — content-comparing uuid-matched rows is deliberately avoided to protect Security Scan purge state (see the pinned test in syncer.test.ts).