Evaluate sum-of-products assignments on raw doubles, boxing once

[lahma]

What

M[i][j] = M1[i][0]*M2[0][j] + M1[i][1]*M2[1][j] + … — the dromaeo-3d-cube MMulti/VMulti kernel
shape — boxed a transient JsNumber per binary node (~7 per matrix cell; 51% of the cube's
allocation in this campaign's Phase-0 census). Assignment right-hand sides that form a left-deep
± chain of two or more products over pure-readable numeric leaves now compute on raw doubles and
box a single result.

Leaves are numeric constants, slot/global identifiers, and one- or two-level dense computed reads
with identifier bases (v[0], m[i][j]).

Safety and semantics

• Every runtime leaf read is pure (slot/validated-global reads, TryGetValueFast dense probes),
  so a decline — leaf not dense, not a Number, TDZ — falls back to generic tree evaluation with no
  observable double effects. A Proxy-trap count test pins exactly-once evaluation on decline.
• Double arithmetic is the spec arithmetic (Number::multiply/add), so integer-valued leaves
  produce identical results; 0 * negative correctly yields −0 (Object.is-pinned).
• The accumulator seeds from the first term rather than +0 — a phantom leading zero would turn a
  (−0) + (−0) sum into +0 (a pin caught this during development) — and only left-deep chains
  arm, so the linear fold reproduces the tree's association, and therefore its floating-point
  rounding, exactly.
• Armed only for all-product chains (≥2 terms), a shape integer-heavy workloads don't produce —
  their dedicated per-op integer paths never see this probe (the constraint from the earlier
  unboxed-islands attempt).

Results

A/B vs main (default job; base side via source checkout since the change was committed):

measurement	before	after
dromaeo-3d-cube driver (±0.7% class)	30.83 ms/iter	26.55 ms/iter	−13.9%
cube allocation census	2.03 MB/iter	1.31 MB/iter	−35.5% (JsNumber share 51% → 18%)
dromaeo-object-array driver (integer canary)	40.42 ms/iter	36.04 ms/iter	favorable/noise
ElementAccess Read / RMW / Append	95.4 / 195.6 / 39.2 ms	93.9 / 187.4 / 38.4 ms	−1.6 / −4.2 / −2.0%
ElementAccess Write / ChainedRead (untouched)	96.7 / 134.7 ms	101.4 / 138.9 ms	+3–5%, inside this family's documented swing; allocations byte-identical on every row

Gating

All-TFM build ✅ · Jint.Tests ×2 ✅ · PublicInterface ×2 ✅ · Test262 99,426/0 ✅

🤖 Generated with Claude Code