Settle the seam a warichu's sublines meet at, and give a furawake its two ends

[P4suta]

Two coordinates a block hides from the line, one at each end of the reading
docs/decisions/stacked-structure-geometry.md
publishes. The first is the sixth constructs variant PR #25 deliberately withheld,
because it was the first shape two of the three engines answered differently from each
other; the second is the furawake half of #24, which needed a structure to be able to say
what it is at each of its two ends.

Fixes #26
Fixes #27

The seam, and which engine was right (#26)

Where a note's two sublines meet, Table 1 states an amount between the character that ends
the first and the character that begins the second. crates/jlreq/src/pipeline.rs offered
that boundary to §3.8.3's ladder; engines/ocaml/lib/pipeline.ml and
engines/racket/compose.rkt did not. Nothing had adjudicated it.

The decision already settles it, and it settles it against the ladder. The published
reading is not "Table 1 states nothing here" — Table 1 states plenty. It is that the line
was not composed from the boundary: the character that ends a subline reports its body
alone and the character after it begins its own row at that row's origin, so the amount at
the seam stands nowhere in any engine's answer. The reading's own sentence for the block's
interior — "a boundary the line was not composed from is not a boundary the line can
adjust, which is why the two of them read the line's geometry and stop at the block" — is
about which object a boundary belongs to, and the seam belongs to the same object the rest
of the block's interior does. §3.8.3 adjusts the spacing of a line; there was no spacing of
the line there to adjust.

The measurement is what makes it more than a preference. A ladder handed a site that cannot
pay divides the arrears over it anyway and recovers only the other share:

※〈〉あ※  half-em note, warichu [3,12), 250 units to give back
             measure   needed   given back   inline_extent
before          3000      250          125   3125  layout.overfull
after           3000      250          250   3000  (none)

An engine that hands back half of what it needs and then reports the line overfull is
contradicting its own diagnostic, which is what tipped this from "two readings" to "one
reading and one defect". The expansion ladder already stopped at the block; the reduction
ladder now asks the same question.

The alternative is publishable and the decision says so: a revision of §3.8.3 that said
what a boundary of a line is, or an Appendix D note saying whether a stage may be offered
a coordinate no amount stands at, would settle it either way. Today the only argument is
the one above, and it is this project's.

A furawake's two ends (#27)

compose.rkt gathers a furawake into one item of the line where a warichu's members stay
items of their own, and two things followed. A column was measured from bare cluster
advances, so a member before a Table 1 amount inside its own column reported 1000 while the
engine placed the next member 1000 units on and charged the line 1500 — the defect #24
fixed for a warichu, at the other structure the same reading is addressed to. And the one
item carried the class of its first cluster, so the boundary after the block was
answered at the wrong end of it, and the amount missing from the column reappeared beside
the structure.

The two are equal and opposite wherever the two amounts are equal, which is why the shape
had to be swept over all 529 class pairs before it could be seen at all:

※〉〈日※ furawake [3,12), two columns, one-fifth-em gap
before  (4500/2200) [... 1000:-600+1000 2000:-600+1000 ...]  ← cancels
after   (4500/2200) [... 1000:-600+1500 2500:-600+1000 ...]  ← Rust and OCaml, byte for byte

※・・日※ the same shape, 500 units inside the column against 250 after the block
before  (4500/2200)   after  (4750/2200)                     ← does not cancel

Both halves had to move together. item now carries a trailing edge — the occurrence a
boundary after the item is read against, which is the item itself for everything one
character wide — and every reading of the before side of a boundary asks for it. A
tate-chu-yoko run carries none: §3.2.5 makes it cl-30 at both of its edges however many
characters it holds.

What changed

• crates/jlreq/src/pipeline.rs: reduction_sites skips a boundary strictly inside a
  warichu or a furawake, the way it already skipped one inside a jidori and the way
  boundary_expansion_site already skipped all of them. Public API unchanged.
• crates/jlreq/tests/public_api.rs: the seam at three measures — the natural width, where
  the amount stands nowhere; the width that needs half of it; and the width that needs all
  of it. It fails on the previous engine at 3,125 against 3,000.
• engines/racket/compose.rkt and spacing.rkt: a furawake's columns are laid out from
  the advances the structure was composed from, and the block carries the character its far
  boundary is read against. across in furawake-item, computed and read by nobody since
  the block's depth moved to block-room, goes with them.
• engines/racket/tests/test-compose.rkt: three cases computed by hand from §3.7.2 and
  Table 1 — the in-column amount and the advance that reaches it, the pair whose two errors
  do not cancel, and the block's own two extents, which neither reading moves. Two of the
  three fail on the previous engine.
• engines/ocaml/probe/census.ml: warichu-pair-inside-reduced, furawake-pair-inside and
  its vertical mirror. Against the previous engines the constructs census reports 15
  differing responses for the Rust one and 434 for the Racket one; against these, none.
• docs/decisions/stacked-structure-geometry.md publishes both readings, why each follows
  from sentences already there, and what a revision would have to say to settle either one
  the other way. engines/ocaml/README.md's census counts were three pull requests stale
  and are now current.

Measurements

Census	Requests	Rust vs OCaml	Rust vs Racket	OCaml vs Racket
spacing	2,116	0	0	0
break	2,116	0	0	0
reduction	3,174	0	0	0
expansion	3,174	0	0	0
vertical	5,290	0	0	0
tate-chu-yoko	4,761	0	0	0
ruby	37,030	0	0	0
constructs	20,102	0	0	0
tabs	31,211	0	0	0
widow	13,225	0	0	0
total	122,199	0	0	0

The Rust sample engine, just conform-ocaml and just conform-racket each answer all
eighty-nine built-in cases at exit 0; raco test engines/racket/tests is 10,196 passing;
just check, just design and just ci are green locally.

No known disagreement between the three engines is left open.

🤖 Generated with Claude Code

https://claude.ai/code/session_01N8KNy2ejPqy81P2hVcRUjs