perf(parser): incremental parse of append-grown session files

[iamtoruk]

CLI half of the desktop-slowness fix: the session cache computed an incremental resume offset for append-grown files (reconcileFile → action: 'appended'), but the parser ignored it and re-parsed from byte 0 — so machines with live agents constantly appending to session JSONL paid a full re-parse of those files on every warm run.

• The Claude JSONL scan now parses only the bytes past the cached offset and merges with the cached turns: continuation turns merge into the last cached turn, boundary re-emits (same msg.id, or the re-read torn line) collapse to the last occurrence keeping the first timestamp (mergeBoundaryCalls, mirroring dedupeStreamingMessageIds), scan-wide dedup keys are re-seeded so cross-file dedup is unchanged.
• Defensive guard in reconcileFile: never resume past current EOF (truncate-then-regrow falls back to a full re-parse).
• No parse-version bump: output is unchanged (only cost), cache shape untouched — a bump would force the very cold re-parse this avoids.

Verified: new integration suite proves the incremental result deep-equals a cold full re-parse (incl. torn-line, inode-change, and offset-beyond-EOF cases) and that the offset path actually activates; full root suite 2178 passed. Real-data check on a 21MB live session file: incremental output byte-identical to cold parse; append-between-runs costs ~0ms extra vs a measurable re-parse penalty before.