Honor explicit start offsets in load

[Metis-dot]

Summary

Honor the documented absolute start argument in load(): when it is explicitly supplied, seek to zero before reading the full content and delegate the decoded-content offset to parse().

Currently, load() reads from the stream's existing position and applies start to that shortened content. This can silently change numeric values, as well as reject a valid document after the stream has been read to EOF.

import io
import json5

fp = io.StringIO('123')
fp.seek(1)
json5.load(fp, start=0)  # before: 23; after: 123

fp = io.StringIO('x 123')
fp.seek(2)
json5.load(fp, start=2)  # before: 3; after: 123

The existing load() docstring explicitly distinguishes start=None from an explicit offset and requires a seekable stream for the latter.

The two-line guard preserves the default: start=None continues reading from the current stream position without calling seek(). Explicit start=0 also follows the documented seek requirement. Negative/out-of-range offset handling and consume_trailing behavior remain delegated to the existing parser.

Verification

• Existing baseline: 75 tests and doctests passed
• Added regression tests fail on unpatched main, including 123 != 23 and 123 != 3
• Patched full suite: 79 tests passed, including CLI subprocess tests against the checkout; doctests passed
• Regression coverage includes text/binary streams, Unicode prefixes, a nonzero initial position, EOF, and default compatibility with non-seekable streams
• Ruff check and format, Pylint, mypy, and git diff --check passed
• Independently verified the published PyPI 0.15.0 wheel and the patch against current main

AI disclosure: An OpenAI coding assistant investigated, authored, and tested this patch, with independent review by another assistant, under the account owner's authorization to contribute to open-source projects.

[dpranke]

Ooh, good catch! I can't believe I didn't write tests for this when I added the feature :(. Nice tests, also :).

[dpranke]

Thanks for the fix!