[9/15] Three copies of the conformability policy, and two of the floor spelling

[mns-nordicals]

What the code looks like now

The conformability contract — a statically known mismatch is a compile
error, dims that cannot be compared statically get a statement-level
runtime guard, provably equal dims need nothing, never warn-and-proceed —
is spelled three times, in three files, by three functions that have to be
kept in agreement by hand:

• guard_conformable_dims() in R/r2f-matrix-blas.R, added by PR 7 for the
  BLAS/LAPACK lowerings. It is the only one that renders a statically known
  extent as the literal rather than size(x).
• emit_elementwise_size_guard() in R/r2f-operators-helpers.R, used by
  maybe_reshape_vector_matrix() for the elementwise operators. Its
  caller does the check-then-emit dance inline at five separate sites.
• An inline per-axis guard inside check_ifelse_branch_shape()
  (R/r2f-conditionals.R), whose private ifelse_axis_verdict()
  re-implements check_elementwise_lengths() — minus the zero-length
  policy, which is why an ifelse() branch of statically known length 0
  slips through where an elementwise operand of the same length is
  rejected.

Two smaller duplications sit alongside:

• check_conformable() (R/r2f-matrix-blas.R) is dims_match()
  (R/r2f-operators-helpers.R) written in list form — same two branches,
  same answers, wrapped in list(ok=, unknown=) so that both call sites
  (solve() routing, the cbind/rbind common-dim check) then have to
  destructure a verdict whose unknown field they immediately collapse
  back to a boolean.
• The stay-in-the-real-domain floor spelling — aint(x) plus a merge()
  correction, written the way it is because Fortran's FLOOR() returns an
  integer and would overflow on a large double — appears verbatim in the
  floor() handler (R/r2f-math.R) and in double %/%
  (R/r2f-arithmetic.R).

Why it matters

• A policy fix has to land three times. Each of the earlier PRs in this
  series tightened the conformability rules, and each had to be applied
  once per emitter. A missed emitter fails silently — it keeps compiling,
  just with the old semantics. The ifelse() zero-length hole above is
  exactly that failure, already in the tree.
• The three emitters have already drifted. They disagree on whether a
  known extent is spelled as a literal or as size(x), on whether a
  zero-length operand is refused, and on whether a NULL hoist is a clean
  error or a crash. None of these differences is intentional.
• check_conformable() invites the wrong reading. Its
  unknown = TRUE result looks like it defers to a runtime guard, but
  neither call site can emit one — a cbind/rbind output dim has to be
  known to declare the variable at all. Both sites turn unknown into the
  same compile error, which is what dims_match() already means.

Expected behavior

No behavior change — this is a maintainability issue. The conformability
policy should be spelled once, in one function, that every enforcement
point calls: the elementwise operators, ifelse(), and the BLAS/LAPACK
lowerings. check_conformable() should be its dims_match() self at both
call sites, and the real-domain floor spelling should live in one place.

The combinatorial conformability grid added earlier in this series
(QUICKR_FULL_GRID=1) is what makes such a change safe to attempt: it
pins values, typeof(), shapes, and the documented compile/runtime errors
for every (mode, shape, operator) cell against plain R.