lgj-abi: hop gathers instead of sweeping — the predicted crossover does not exist - #40

…es not exist

Closes ISS-LGJ-HOP-SWEEPS-FULL-POPULATION. The issue named a gather as the
next rung and said its merit depended on a DENSITY CROSSOVER -- sparse
frontiers favouring a gather, dense ones favouring the sweep's sequential
vectorised access -- and that this had to be measured rather than assumed.

Measured. THERE IS NO CROSSOVER.

`examples/hop_gather_vs_sweep.rs` runs both shapes over five populations
(1_024 ... 262_144) x twelve frontier densities, asserting BYTE-IDENTICAL
output at every one of the 60 configurations plus an anti-vacuity guard
against an empty hop (which would let both shapes "agree" on nothing and time
two empty loops). Gather wins all 60:

  n_rows      0.01%      1%     25%    100%
   1_024       361x    113x    7.3x   1.73x
   4_096     1_295x    143x    5.9x   1.75x
  65_536     2_666x    117x    2.5x   1.78x
 262_144     2_754x     90x    2.6x   1.73x

WHY IT HOLDS EVEN AT 100% DENSITY -- the part the prediction got wrong. A
sweep MATERIALISES an n-element per-row intermediate that each row reads
exactly once, so its cost is never amortised. At full density it does
everything the gather does PLUS allocate, zero, write and re-read n u32s. It
is strictly more work at EVERY density, not merely at sparse ones. The
prediction reasoned about access PATTERN and missed that one shape simply does
MORE.

Consequence for the design, and it is the better outcome: NO THRESHOLD, NO
DISPATCH, NO HEURISTIC GATE. A real crossover would have required one, with
the two-sided evidence such a gate demands. Its absence makes the change
unconditional.

A SCALAR GATHER BEATS A VECTORISED ndarray::simd SWEEP. Same lesson the
object-model ladder taught: the win is in not doing the work, not in the
vector width. No SIMD primitive is called in the hop now and none is missing
-- `simd_rowstore_facet_match` remains the kernel behind `lgj_row_facet_match`
and `simd_rowstore_classid_mask` behind `lgj_op_eq_classid`, so neither is
orphaned (checked, not assumed).

END TO END through Component G, the independent instrument:

  native_hop         1%/4096   1%/65536   25%/4096   25%/65536
  original (32x)       479.0    24798.3      521.2     23633.9   us/op
  one-pass (#39)       374.7     7120.3      375.8      8076.9
  gather                 1.9       34.4       61.0      3534.1
  vs original           246x       720x       8.5x        6.7x

The ordering the component was built to test has INVERTED: native is now
fastest at every configuration, 2.0x-13x ahead of the best scalar arm, having
begun this arc slowest at every configuration by 2.6x-165x.

The two instruments are independent and agree: the Rust probe measures the
sweep at 6754-8118 us at 1%/65536 against JMH's 7120 us.

HONESTY on the JMH re-run: it was noisier than the previous one --
`classidScan` at 1%/4096 reports 25.6 +/- 61.9 us, an error bar larger than
the score -- so the SCALAR absolutes in that table are weak. Native's own
errors are tight (1.9 +/- 0.3, 34.4 +/- 8.1) and a 720x change is far outside
any noise this container produces.

ALSO FIXES A DOC I STALED ONE PR AGO. `lgj_hop`'s "Kernel composition" block
still described the pre-#39 design (one full-width classid sweep PER FACET, a
scratch buffer "REUSED across every participating facet"). #39 changed the
code and left the prose. Rewritten to describe the gather, with both
superseded shapes named so the arc is legible.

The overflow guard is KEPT as an explicit check rather than dropped or hidden
behind an `_n` binding: the gather no longer needs `n`, but row indices are
still cast to usize, so a store whose row count does not fit must still be
refused there.

STILL OPEN, named rather than hidden: the one shape that could favour a
precomputed per-row mask is REUSE -- memoising it across many hops on the same
(store, classid). That is a caching design with its own invalidation
questions and is deliberately not this function's.

Gates: lgj-abi 134/134; AllTests 304, GraphHopTest 66 (incl. G3 at its
unchanged allocation floor), TradesParity 12, TradesAllocation 3, BricksAuth
62 = 447 Java checks; clippy -D warnings and fmt clean. No ABI change, no
signature change, no minor bump -- an internal shape change behind an
unchanged contract.

Board artifacts same commit: ISSUES resolved with the corrected mechanism and
the remaining reuse question, RESULTS.md § G carries both tables and the
noise caveat, raw probe output banked at
.claude/board/hop-gather-vs-sweep-crossover.txt.