[gspark-etri]

Completes native CCGrid E6/E7/E8 experiments on RNGD SDK2026.4, using one card and raw-linked SHA256 evidence. All18 condition groups have n=5, with median/IQR and exact independently fresh target-token checks. The prepared-serving integration branch is the base; main remains outside this change.

• E6: three-arm smoke +30 main samples: same B4/S256→B4/S512 switch using full pull+restart, delta apply+restart, resident reuse; node chunk cache empty/prepopulated separately. Actual Pod→first-token and monotonic trigger→first-token clocks, phases and byte evidence are retained. Results/conditions/figures, all statistics and raw links.
• E7: four smoke +30 main samples: sparse projection and LM fixtures, inplace rows / entire touched stacked-group replacement / sequential worker restart. Update→first-token medians: rows0.632–0.635s, group replacement0.963–1.127s, restart19.948–20.359s. Actual PR76 v1 CLI inputs, all fixture file SHA256, fresh refs, raw native counters and actual Job policy accompany the measurements. Results, statistics/raw.
• **E8:**30 native samples, three directed B/S switches × actual first-use/prepared states ×n5. All113 weight buffers and preloaded modules retained; no second shared-weight upload. Each target equals fresh load of that same variant. The superseded B8-vs-B4 failure remains an observation (lane0/token4) with its original record. Results, statistics/raw.

The strict onggi.switchplan.v1 parser retains manifest/source metadata, validates complete partitions/classification totals and paths, rejects unknown schemas, and keeps legacy compatibility. This is a versioned input consumer, not native B2 conformance. Fit assessment records missing actual artifact-space device readback, retained rollback, owner/drain protocol and existing private SDK dependencies. Runtime follow-up #39; neutral fixture work onggi#79. Separate onggi PR80 fixes missing-unit verification and has12 regressions plus green Go/S3/fake-kit checks; fake-kit success is not native B2.

Validation: immutable executed sources E6=530195d, E7=f0f08bf, E8=bbceb17. Root independently verified original exports, frozen Git source blobs, normalized numeric fields/tokens/identities, canonical Git raw hashes for all90 main samples and all three documented CLI summary reproductions. All native output type/allocated-index gates pass; compile/retry0. Local full suite590passed8skip4subtests; final publication PR CI41054624436 read596passed2skip4subtests. One cdx- Job/card at a time with request/limit1, TTL/deadline and UID cleanup; final actual API shows no owned Job/Pod or active RNGD requester.

Limits: native decode, not HTTP gateway. E6 uses colocated POSIX delivery; empty/prepopulated refer to node chunk cache, OS page cache/background load uncontrolled. Priming/base setup count in Pod time; bootstrap scope is explicit. Upload timers combine host layout and H2D; logical bytes are not physical DMA. Pod clock skew bound unknown despite synchronized status. E7/E8 have no artifact-delivery or Pod timings and are not pooled with E6. Token-ID parity does not establish dense-logit bitwise equality. Paper NUMBERS.md unchanged; optional ATOM #37 deferred.

Completes experiments #34, #35, #36; B2 remains open in #39.