Fix _stream line-offset overstatement for chunk-spanning lines

[valiantone]

Summary

Fix the line-offset arithmetic bug in JSONLInspector._stream discovered by the native helper spike (#48, bench in #84).

Bug

When a line spans the 1 MiB read-chunk boundary, _stream computes line_start = offset + pos in (carry+chunk) coordinates (src/hotmem/inspectors/jsonl_inspector.py:129), overstating file offsets by len(carry) for every subsequent line.

Evidence on the spike corpus (events_10mb.jsonl): the malformed line sits at byte 6,292,320 (manifest-committed); the real _stream reports 6,292,650 — off by +330, exactly the carry length at the crossing. The spike's C scanner reports the correct offset.

Affected outputs: unsupported_reason offsets and byte_ranges (both feed FileInspection → API/MCP/CLI).

Fix

line_start = base + pos where base = offset - len(carry) — the replica arithmetic in bench/native_spike/py_baseline.py::py_scan_only, which achieves full parity with the C scanner.

Acceptance criteria

• One-line arithmetic fix in _stream.
• Regression test with a fixture whose early line spans the 1 MiB chunk boundary (the spike's fixtures are reusable), asserting byte_ranges and unsupported_reason offsets against a plain full-file read.
• All existing inspector tests pass.

Context

• Spike: Native helper spike: Rust, C, or WebAssembly fast-path primitives #48 · bench PR: bench: native helper spike — C/WASM vs Python fast-path primitives (#48) #84 · bench/native_spike/README.md ("Discovered: _stream line-offset bug")