perf(indexer): incremental byte-offset JSONL parse — 604x on appends (v1.1.90)

The session indexer re-read and re-parsed the WHOLE transcript on every
append (a ~4 KB append to a 211 MB live session = a ~525 ms re-parse plus
rewriting every archive + FTS row). Now it reads only what changed.

How (history.rs::index_adapter_session):
- cc_sessions gains last_indexed_byte_offset + last_indexed_line_count
  (additive migration, NOT NULL DEFAULT 0 — legacy sessions full-parse once
  to set the cursor, then go incremental).
- when an indexed file only grew, seek to the saved offset, parse the tail
  up to the last newline (half-flushed events are never indexed), and merge
  deltas: append archive rows with continued message_index/source_line, bump
  day buckets, sum token totals, recompute cost from the new totals.
- shrunk/rotated file (file_size < offset) falls back to a clean full reparse.
- parse_claude_session/parse_codex_session are now thin wrappers over one
  generic adapter path.

Proven:
- incremental_index_matches_full_reindex_byte_for_byte — an incremental
  index is byte-identical to a one-shot full re-index (totals, cursor, cost,
  every archive row, day buckets).
- file_shrink_falls_back_to_full_reparse — rotation rebuilds cleanly.
- bench_incremental_reindex_vs_full (23.5 MB): full reparse 1275.9 ms →
  incremental append 2.1 ms = 604x; gap widens with size. (docs/PERFORMANCE.md)

215 tests pass. No language rewrite — this was algorithmic, exactly as the
harness predicted.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: sarthakagrawal927

apps/desktop/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@code-reviewer/desktop",
-  "version": "1.1.89",
+  "version": "1.1.90",
   "private": true,
   "scripts": {
     "dev": "lsof -ti:1420 | xargs kill -9 2>/dev/null; vite",