docs(agents): record §5 of the conventions spec, and the markdown line-width call - #388

Two conventions that were live but unwritten.

**Stage contract docstrings (§5).** The 2026-07-11 v2 implementation
conventions spec fixed nine sections before the first line of code.
Its own header attests to distilling "§1–§4, §6–§9" into this file on
2026-07-12 — a range that reads as exhaustive unless you notice §5
missing from the middle of it. §5 is the one section that never
reached a committed home, and the 2026-08-16 spec-residue sweep
(#381) did not catch it, most likely for the same reason.

It has been followed anyway: all nine pipeline stages carry the
Consumes:/Produces:/Reads: docstring, and the two _pipeline/ modules
that are not stages say so in their first line instead. So the
convention was load-bearing and unenforceable at once — a new stage
could have omitted it and no reviewer would have had grounds. Now
recorded, including that the absence of the three lines is itself a
claim about the module.

**Markdown line width.** Measured rather than argued: wrapped and
unwrapped markdown render identically, docs/design/ is not in the
Sphinx build, and unwrapping is byte-identical since newline→space is
1:1 — so no rendering, build, size or token difference exists. The
cost is one-sided, at 74% of the deleted words in #386's
decisions.md diff being reflow rather than change. New prose goes
unwrapped; existing paragraphs stay until rewritten, because a bulk
rewrap resets git blame across 31 and 43 authoring commits in
rules.md and decisions.md, and that provenance is what dates a stale
claim.

The exception is enforced, not stylistic: rules.md's example lines
and pointer lines are parsed per line, and a wrapped example is a
hard error. Statement prose is free because the citation check
normalizes whitespace first.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The note telling you not to hard-wrap prose was hard-wrapped across
two paragraphs and 19 lines, having been offered as one. Now one
unwrapped line, which is both the convention it states and the length
it promised.

Dropped in the trim: the per-medium enumeration of what makes no
difference (rendering, build, size, tokens), the blame-commit counts,
the .git-blame-ignore-revs escape hatch, and the exception's field
list. The measurements stay in this PR's description and #386's
history; a convention needs the rule and the one number that justifies
it, not the workings.

Note the section it sits in was already wrapped while "2.0 API
modules" below is not, so the file was mixed before either note
existed -- which is why the §5 bullet in the previous commit is
unwrapped and correct without anyone deciding.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Three errors in the §5 bullet, all found in review.

It said stages are "named exactly as rules.md and the specs name the
stage". rules.md cannot be that authority and says so in its own
preamble -- "Statements are implementation-free: no stage names, no
function names, no regexes" -- so the bullet asked rules.md to do the
one thing it forbids itself. Nor does `implemented:` supply it: that
field names the modules honoring a rule, and a rule can list several
while a module honors several, so it carries no stage naming at all.
The committed authority is `_pipeline/__init__.py`'s STAGES for the
set and ParseState's docstring for the field-ownership map, pinned by
tests/v2/pipeline/test_state.py. The original §5 pointed at the v2
core API spec, which is gitignored -- which is why the distillation
needed a different anchor rather than a transcribed one.

It said "all nine stages carry it". There are eight; STAGES calls
itself the eight-stage fold. The ninth module is _assemble, which the
same sentence describes as not a stage -- the count contradicted its
own clause.

Trimmed with it: the spec-header attestation story, which is why this
section was missing rather than what the convention is, and belongs in
#388's description.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>