v2 spec: the pre-init refusal (precision) and chs_preview_row reads exactly one row (prose, fp unchanged)

[EricAndrechek]

Two prose statements in the v2 spec, made before the artifact producer pins the header:

1. Teardown (a precision fix to v2 spec: a declined export's out_export is NULL; calls before chs_initialize refuse (prose, fp unchanged) #621): the pre-init refusal excepts chs_initialize alone (chs_shutdown returns void, so it was never among the chs_status calls); and after chs_shutdown has run, every such call refuses the same way (CHS_INVALID_ARGUMENT, ClickHouse code 0, a message saying chs_shutdown has run).
2. chs_preview_row reads the whole body, which holds exactly one row (the producer's requirement for its next library build): the body is read with chs_preview_batch's reader; a malformed tail or trailing content the reader does not skip is rejected with the reader's own code, as a server's INSERT rejects it, while what the reader skips (a trailing newline) is fine; zero rows or more than one is CHS_INVALID_ARGUMENT, naming the call, with chs_preview_batch the call for many rows; and the row document equals the single row of chs_preview_batch's document for the same body. Published builds up to and including 20261011.105019 answer from the first row and ignore the rest.

Prose only (spec/abi-v2/docs.md, regenerated include/v2/chtypes.h and docs/reference/abi-v2.md). Both statuses are already in chs_preview_row's may_return. Measured: gen.py --major 2 --check passes, the fingerprint is unchanged (d60a681e…), and the header sha256 is b2d9b8290ffc311a57f91998495ae0588a0f93eeb6bb8b3efc5efb549f089315. markdownlint, dprint, lint-spelling, lint-cited-paths and lint-public pass.

🤖 Generated with Claude Code