Preserve runs of compact assignment rows

[jserv]

A comment that documents a mapping one row at a time, the ir->imm = lui immediate shape a decoder or a register table is written in, was packed into a paragraph like prose, which loses the mapping. Those rows now classify as a preformatted kind and go out verbatim. What decides a row is the left-hand side, which must be a single code-like token, so a sentence that merely carries an equals sign stays prose; the operator may be bare or compound, spaced or not, and alignment padding before it is layout rather than a disqualifier, since column-aligned tables are the common spelling. A run of two or more adjacent rows freezes whole and whatever follows is simply the next paragraph, so a table sitting directly above the sentence that explains it survives intact, while a lone count = zero when the queue drained is a sentence and still wraps.

Run membership is decided after every other check, over the lines nothing else claimed, so a run cannot reach through a code sample, a table, a fence or a verbatim block. Deciding it on how a line looks in the source instead let an indented ret = foo(a, b); pair with the sentence under it and freeze that sentence alone, which is the case the run rule exists to reflow and a common shape in C headers. Two rules this leans on needed work: art declines a mapping row, since = is in the art alphabet and a short row cleared the density threshold on the very characters that make it a row, except where the line is rule-decorated, or the packer builds a line ruled on both ends that the next pass strips; and the merge gate reads through a bookend, because strip_decorative_bookends runs between the two stages.

The first commit is independent and stands on its own against the base branch: line 0 of a /* <content> block was excluded from canonical re-emit and took raw replay, so it kept a decoration its siblings had lost and the pass never converged. That affects label and banner rows with no assignment row involved.

Verified against the pre-change binary over 900 system headers and Rust sources, about 550k lines, at column limits 40, 60, 80 and 100. Five files differ at 80 and each is a mapping table the old binary had run together: the adler32 derivation in adler2, the write defaults in brlapi.h, the size definitions in nettle/ccm.h, the permission rows in xenstore.h, and two lines of literal example code in webp/decode.h. Preserved rows match their source bytes including the alignment indent, no comment word moves in any file, and the set of files whose second pass differs is unchanged. A 4000-case fuzz over an operator-heavy vocabulary and a 1500-case fuzz mixing multibyte characters with every operator byte, both run against the base binary for comparison, found no new non-convergence and no panics. Run handling is linear, measured at 3000, 10000 and 20000 rows in one comment, and the branch costs about 0.5% more instructions on real headers. cargo test is 452 green, with cargo clippy --all-targets and cargo fmt --check clean, and commentflow --check on the changed files matches the base branch.

Left as it stands: a run of two or more overlong rows keeps sitting over the column limit, the same way a table row or an indented code sample does, because a real mapping row cannot be wrapped without destroying the mapping. That boundary is pinned by a test rather than fixed. Two pre-existing issues found during validation are untouched here: a /*- opener merge that does not converge, visible in two lz4 headers, and a quadratic prefix rescan in is_file_start_prefix that a one-line conjunct reorder would fix, measured at 5% of instructions on real headers and 53% on comment-dense input.