MLA-SHARED-1: the four-family map earned at both banks — MLA absorbed, shared experts refused on economics

[chrishayuk]

What

The fourth byte family joins the representation vocabulary, and the fifth is
refused on economics. MLA projections and shared-expert branches become
encodable scopes; the composed four-family map is judged at 8,192 positions on
both banks under kimi-logit-balanced-v1 and benchmarked.

Five families make up the decoder's BF16 read per token — routed experts 49.2 %,
KDA projections 25.2 %, output head 12.6 %, MLA projections 6.8 %, shared
experts 6.1 %. Four are now admitted under one frozen contract.

The engine change

• MLA encoding threading — the fourth verbatim application of the pattern:
  MlaDeviceWeights/DeviceAttn::Mla carry a projection encoding (four wide
  projections only; kv_a_norm stays f32, the latent cache always holds decoded
  values); all three MLA grouped dispatches plus the at-offset cache write select
  their kernel by encoding.
• Shared experts needed zero kernel work — their dispatch has been
  encoding-aware since the expert rung; only the bytes were BF16.
• The bench harness learned the same scopes — it timed KDA and the head but
  could not see MLA or shared, so a four-family map could be judged for quality
  and not for speed. It now shares one scope-parsing surface with the quality
  probe, and a head-only map no longer silently times two identical arms.
• The report carries the contract it is judged by — both the strict v3
  verdict and the balanced-v1 verdict travel in every report, so a claim of
  admissibility no longer has to be re-derived from the bank by hand.

The measurement

MLA depth curve — the fourth family with the same shape every other family showed:

L23 5.75e-5  |  L26 1.24e-4  (plateau)
L19 8.43e-4  (the 15x jump)  |  L15 1.06e-3  |  L11 3.70e-3
all 7 MLA layers  2.61e-2  — 450x the L23 cell

The tower's token-distance curve ACCUMULATES (2.6e-4 → 2.9e-3) where the late
cells' does not. Scope chosen by the jump rule: MLA{23,26}, non-contiguous
by design.

Shared experts refused on economics, not capability. Their dispatch has been
encoding-aware since the expert rung, so admitting them costs no kernel work:

scope	cost	bytes saved	cost per MB
MLA L23	5.75e-5	27.3 MB	2.1e-6
shared L25	1.240e-3	6.6 MB	1.9e-4

~90x worse per byte. The pair {24,25} even composes sub-additively (1.09e-3 vs
1.24e-3 solo) and still loses. The first refusal in this programme that is an
optimizer's judgement rather than an engine limitation.

Authority — both banks, 8,192 positions, kimi-logit-balanced-v1: PASS, no
failed criterion.

criterion	limit	three-family	four-family	budget
selection kl p99	3.5e-3	2.4006e-3	2.3791e-3	68.0 %
top-1 mass displaced max	0.12	9.415e-2	5.546e-2	46.2 %
top-10 mass p99	0.12	7.134e-2	6.465e-2	53.9 %
route mixture p99 / max	0.15 / 0.25	0.1258 / 0.1993	0.1248 / 0.1993	83.2 / 79.7 %
held-out kl p99	3.5e-3	2.6378e-3	2.6174e-3	74.8 %
top-1 mass displaced max	0.12	5.8313e-2	5.8313e-2	48.6 %
route mixture p99 / max	0.15 / 0.25	0.1262 / 0.2099	0.1257 / 0.2099	83.8 / 84.0 %

MLA{23,26} is behaviourally absorbed. On the held-out bank both max
statistics are IDENTICAL to the three-family map to four decimals — the worst
single overturn and the worst single routing displacement in 8,192 positions are
the same events, unchanged, after adding a family. Counts move ≤ 2 %. On the
selection bank the worst top-1 severity fell 0.094 → 0.055 while the flip count
went 130 → 129: one position reshuffling, not an improvement, and not claimed as
one.

The route budget is the binding constraint (83-84 % of limit against KL's
68-75 %). It is the only criterion measuring a discrete consequence — which
experts got selected — so it does not average out over 8,192 positions the way
KL and mass displacement do.

Decode — the cost model scored against the machine. Prediction registered
BEFORE the run from the byte ledger alone: 957 MB of the 5.985 GB per-token BF16
read removed (16.0 %), predicted 1.15-1.16x.

session	baseline	candidate	wall	GPU ms/token	GPU speedup
1	35.81 tok/s	40.79 tok/s	1.139x	26.87 → 23.43	1.147x
2	35.16 tok/s	40.52 tok/s	1.153x	26.97 → 23.56	1.145x

Measured 1.146x GPU, 1.139-1.153x wall. The prediction was accurate to ~1 %
but on the optimistic side — the GPU measure sits just below the band, not
inside it. Score the GPU column: it is stable to 0.05 ms across blocks while
wall carries 1-6.6 ms/token of intermittent non-GPU overhead. Conversion:
16.0 % of bytes removed → 12.7 % of GPU time removed, 0.79-0.80, against a
1.190x pure-bandwidth roofline.

Instrument hazards catalogued

Two flat-instrument episodes were diagnosed and are frozen as evidence under
kimi-quality-bank-provenance/QUARANTINE.md:

signature	reading
kl_p99 exactly 0.0, coverage exactly TOP_N/vocab	the GPU completed command buffers without executing them
coverage far under the bank's floor, top10_mass_displaced p99 exactly 1.0	the instrument degraded PART WAY THROUGH a ~25 minute run

The CoveredMass criterion refused both — a run that could only "pass" by being
blind is exactly what it exists to catch. Root cause: the probe held a third
full stack (null_partner, ~17 GB of wired attention banks) needed only by the
2-minute null arm through the entire measurement. It is now dropped immediately
after the null arm, and the authority chain checks instrument health after
EVERY stage rather than only up front.

Gates

All larql-vindex CI gates green, verified on an ISOLATED worktree holding HEAD
plus only this diff — the shared worktree currently carries another in-flight
change whose modules are not yet declared, so a gate run there would not have
been attributable:

fmt · check --all-targets · check --examples · check --features gpu --all-targets
clippy --all-targets -D warnings · e0_generation_boundary (10)
cargo test -p larql-vindex (3418 lib + all integration binaries)
benches compile · gpu lib tests (3469 passed, 4 ignored)

Note for a follow-up, not fixed here: cargo clippy --features gpu — which CI does
NOT run — reports 4 pre-existing errors in kimi_moe_metal/grouped/represent/.
GPU-gated test code is currently unlinted by CI. A fifth, in the file this PR
touches (a condition duplicated with is_empty()), is fixed.

What this opens

Four byte families admissible under one frozen contract; the fifth declined on
bytes-per-behavioural-cost. Q8_0 was the INSTRUMENT that mapped the topology,
not the destination — every family showed the same late-plateau / interior-cliff
shape, which is what makes a search tractable. The manual phase ends here.
BALANCED-SEARCH-2 optimizes codec choice under the vector of behavioural
constraints, not a scalar bytes-per-KL, with the full-bank composed gate
authoritative.