fix(claude): tolerate duplicate snake/camel keys in session JSONL

[tuchg]

Fixes #45
Fixes #46

Problem

Claude Code v2.x (v2.1.206, v2.1.208, …) writes both session_id and sessionId (and other snake/camel pairs) on the same JSONL line. serde's generated struct visitors reject the second occurrence as a duplicate field, so every assistant/user line fails to parse — Lucarne sees no Claude activity at all (probe, history, and watch all fail with duplicate field session_id).

Fix

Wrap all claude JSONL parsing (probe, history body, watch meta, plus nested content/usage/progress parses) in a DupTolerant serde Deserializer:

• Buffers each object as borrowed (key, raw value) pairs — zero-copy, no allocation for canonical keys
• Folds camelCase spellings onto their snake_case field via a generic camel→snake canonicalization (works for every current alias and future ones, including fields with no declared alias)
• Drops repeats, last occurrence wins (same semantics as serde_json::Value)
• Replays the deduplicated object; nested structs (e.g. progress payloads) are covered recursively

All logic stays inside the claude provider — no provider ids, cursor formats, or parse rules moved into shared layers (per AGENTS.md boundaries).

Tests

• parse_claude_json_str_tolerates_duplicate_keys_last_wins — dual session_id/sessionId + message_id/messageId lines, asserts last-wins
• probe_session_meta_accepts_dual_key_lines — probe path on the exact bug shape
• reader_accepts_dual_key_lines_and_nested_progress — watch/history path with dual keys + nested fullOutput/full_output progress payload
• All existing 219 agent-sessions tests pass; clippy clean for the touched file

Tradeoff

claude_parse bench: 37ms → 60ms per full session (~1.6x). Tolerance requires a second pass over the line — streaming serde deserializers can't rewind to dedup in one pass. Watch deltas are single lines (µs-scale), so the practical impact is negligible.