kolibri1 TT: three eager-path levers — decode step 1077-1254 ms to 610 ms (-47%)

[lu-zero]

The captured-decode investigation (docs/gap-kolibri1-tt-captured-decode.md) concluded full trace capture is NOT feasible under ~3 days for kolibri1 — the MoE routing is host-side and data-dependent (per-layer full-tail router-logits download + streaming expert fetch with per-miss D2H readback/verify/CPU dequant), and ttnn capture is fatal across those round-trips. The missing Warm*/graph scaffold is ordinary porting but worth ~zero while the routing serializes every layer on the host. In its place, this lands the re-ranked EAGER levers, measured with 3-run medians against the bench's measured +-10% noise:

1. Decode bench cache allocation at block_size 32 (bench only; the correctness gate keeps 16): the %32 TILE guards now pass and decode reaches median 953 ms/step (-17%, all three runs below the best baseline). The win is host-path geometry — the device attention arms still decline (the raw host NHD cache carries no rank-4 device KV shadow), recorded honestly.
2. The streaming expert path's SECOND per-miss D2H readback (memo dequant) measured 60.5 s cumulative over a gate run — ~315 ms of every ~940 ms step. Fix: memo dequant reads the slot shadow's already-verified host image (bytes identical by construction; readbacks -> 0). Decode: median 610 ms/step (-36% vs lever-1 base, -47% vs baseline).
3. A verified-bitmap fill-verify skip: honest no-op (0 skips fired — every fill is first-touch after eviction), kept env-gated off the critical path.
4. Lever 2's full port scope (device-side routing via a vt::MoeRouterTopK TT port, the prerequisite for trace capture later) is documented in the gap doc rather than ported here.

A router-bias host-cache prototype from the investigation is kept (inert to numerics) but recorded as NOT a perf result: removing 50 per-layer blocking syncs moved nothing — those wait on the tiny router GEMM tail, not the queue; the serialization lives in the full-tail router-logits download and the expert-fetch round-trips.

Also recorded honestly: the greedy chains are NOT reproducible run-to-run in any arm (consistent with the B2b-ii record's 7-flip residue, adjudication pending #3438-style decision). Indicative throughput after the levers: ~1.6 tok/s decode (from ~0.93), prefill walk 149-194 -> 105-119 s.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:zai/glm-5.3-flash [maki]