diff --git a/README.md b/README.md
index 81b116a..e909d3d 100644
--- a/README.md
+++ b/README.md
@@ -1,6 +1,6 @@
 # Frontier MoE on consumer Blackwell (SM120)
 
-**Official vLLM v0.24.0 + a 7.4k‑line patch** that serves frontier Mixture‑of‑Experts models —
+**Official vLLM v0.24.0 + a 14.0k-line, 70-file patch** that serves frontier Mixture-of-Experts models —
 **GLM‑5.2 (753B)**, **DeepSeek‑V4‑Flash (159B)** and **Kimi‑K2.7‑Code (1T)** — on
 consumer/workstation Blackwell (RTX PRO 6000, RTX 5090), hardware their official checkpoints
 cannot even fit on. Three ideas carry it:
@@ -10,7 +10,8 @@ cannot even fit on. Three ideas carry it:
    gate) restores precision exactly where it matters.
 2. **Tiered expert residency** — when even the 2‑bit base outgrows VRAM, it moves to pinned
    host RAM — and, one tier further, to an **NVMe pack file with a pinned‑RAM arena** — and
-   the GPU becomes an **expert cache** (miss → batched fetch + bit‑identical graph replay).
+   the GPU becomes an **expert cache** (miss -> batched fetch + graph replay, with explicit
+   telemetry for any accepted second-order approximation).
    That puts 753B on two 96 GB cards and 159B on a single RTX 5090, and the packs double as
    a **persistent quantization cache** (reboots skip the re‑quant).
 3. **A rebuilt serving base** — vLLM v0.24.0 actually working on SM120 (the release is
@@ -63,10 +64,13 @@ medians; prefill = 8k‑token prompt, uncached):
 | 2× RTX PRO 6000 (TP2) | 210 tok/s | 5 790 tok/s | 512K | — |
 | 4× RTX 5090 (TP4) | 214 tok/s | 6 100 tok/s | 16K | — |
 | **1× RTX 5090 (32 GB)** | **~31 tok/s** (14 GiB pool + NVMe stores) | ~400–540 tok/s | **32K** | **~30 GiB** |
+| **1× RTX 5090 (32 GB), no-MTP quality config** | not measured | not measured | **128K; PASS at 120K** | **≤100 GiB eval cgroup incl. cache** |
 
 Retrieval behind the window column: needle PASS at 453K on the PRO 6000 (947K‑token KV
-measured) and at 29.7K on the single 5090 (131K‑token KV). "—" in host RAM = all‑VRAM
-config, no host expert store.
+measured), at 29.7K on the single-5090 throughput configuration (131K-token KV), and 3/3 exact
+at 120K on the canonical no-MTP 128K quality configuration (720,677-token KV). The 128K row is
+a quality/context result, not a throughput measurement. "—" in host RAM = all-VRAM config, no
+host expert store.
 
 **Batched serving** (aggregate decode tok/s at N concurrent streams; per‑stream in
 parentheses at N=32):
@@ -81,6 +85,10 @@ configs. MTP also runs under **pipeline parallelism** (draft propagation + draft
 share across ranks): DS4 on 4× RTX 5090 **PP4** does 184 tok/s vs 93 without (~2×), and greedy
 decode under PP is **bit‑deterministic** (6/6 identical runs, with and without MTP).
 Methodology: **[docs/v024-port.md](docs/v024-port.md)**.
+The guarded single-5090 W2 quality/context result, including P0/P1/P2 evidence boundaries, is
+reported in **[docs/benchmarks/ds4-w2-5090-2026-07-11.md](docs/benchmarks/ds4-w2-5090-2026-07-11.md)**;
+the sanitized receipt index is
+**[evidence/public/ds4-w2-2026-07-11/](evidence/public/ds4-w2-2026-07-11/)**.
 
 ---
 
@@ -128,11 +136,12 @@ routed working set). MoE routing is concentrated enough to make this practical:
 coverage serves ~96% of token→expert routings** on DS4, **~51% serves ~91%** on GLM —
 measured live, not simulated.
 
-Misses stay correct through the gate's replay trick: the desc kernel zeroes a missing
+First-order misses are restored through the gate's replay trick: the desc kernel zeroes a missing
 expert's contribution and bumps an in‑graph miss counter; the runner fetches **all** missing
 routed experts in one batched pinned‑H2D transfer (51.6 GiB/s here; a 64‑expert fetch ≈ 3 ms)
-and replays the step's graph once — **bit‑identical** to a fully resident forward
-(unit‑tested). A **miss‑tolerance knob** (`VLLM_MOE_W2_BASE_MISS_TOL=k`, runtime‑tunable)
+and replays the step's graph once. The result is bit-identical to a fully resident forward
+only when that replay creates no second-order misses (unit-tested for the zero-residue case).
+A **miss-tolerance knob** (`VLLM_MOE_W2_BASE_MISS_TOL=k`, runtime‑tunable)
 skips the replay when ≤ k of the step's ~600 routings miss — +12% decode on GLM TP2 at
 tol 8 (28.3 → 31.7 tok/s) with clean quality probes (arithmetic, PL coherence, needle
 retrieval; quantitative eval pending).
@@ -149,14 +158,20 @@ e.g. `--gpu-memory-utilization 0.95`) before touching any other knob.
 **Misses are restored adaptively.** A replayed step can re‑route onto experts the first
 pass never fetched (second‑order misses); the runner re‑checks after each replay and keeps
 replaying **only while the step is within `VLLM_MOE_W2_FP_THRESH` of miss‑free** (default 0
-= the mandatory first‑order restore only, the throughput‑optimal setting; raising it buys
-bit‑deterministic fixed points on converged working sets at a decode cost — runtime‑tunable
-via `VLLM_MOE_W2_FP_THRESH_FILE`). The accepted residue is second‑order only and
-KPI‑visible (`fp-residue`).
+= the mandatory first‑order restore only, the throughput‑optimal setting; raising it can chase
+zero-residue fixed points on converged working sets at a decode cost, but does not itself promise
+deterministic output — runtime‑tunable via `VLLM_MOE_W2_FP_THRESH_FILE`). At the default
+threshold, newly fetched second-order
+experts are available to later steps but their current-step contributions remain zero. This
+is an explicit throughput/quality approximation, not a miss-free or bit-deterministic claim;
+its frequency and pair count are KPI-visible as `fp-residue` and are reported beside quality.
 
 Results: **DeepSeek‑V4‑Flash 159B on one RTX 5090** (72.7 GiB of 2‑bit planes vs 32 GB of
 VRAM): ~31 tok/s steady with MTP, 32K window served, coherent — and ~30 GiB of host RAM
 with the NVMe stores (below, RSS‑measured) instead of ~80 GiB pinned.
+The canonical no-MTP quality configuration separately serves a 128K window: 120/120
+machine-exact and semantically correct, 0/120 frozen-rule sink detections, and 3/3 exact
+retrieval at 120K prompt tokens. Throughput was not measured for that configuration.
 **GLM‑5.2 753B on two RTX PRO 6000**: 28–32 tok/s with the full three‑tier stack (NVMe 2‑bit
 base + pinned arena → GPU 2‑bit cache → GPU FP4) at a 128K single‑user window — see the GLM
 table above. Neither model can otherwise run on that hardware at any precision.
@@ -214,15 +229,53 @@ Operational notes:
   drive on a PCIe **Gen3 x4** link (3.7 GB/s). Cold working‑set shifts and first‑touch
   prefills do pay drive speed. `VLLM_MOE_W2_TIER_DIRECT=1` switches misses to O_DIRECT
   (hard RAM budget, page cache stays flat; raw drive latency on every miss).
-- **Prefill can't wipe the arena** (scan discipline: prefill working sets fill free slots
-  but never evict the decode hot set), and the arena's hot set persists to
+- **Cold pack rebuilds fail closed instead of consuming the host.** MXFP4 expert parameters are
+  zero-sized at model construction and materialized only for the layer currently loading; that
+  layer is quantized, durably written, and released before the next one starts. Exact pack
+  identity skips the build entirely, and completion guards prevent a streamed layer from being
+  built twice. While W2 + a pack store are active, the sequential safetensors loader also
+  preflights every shard, requests eviction of
+  each released shard with `POSIX_FADV_DONTNEED`, and retries after the model consumer unwinds;
+  the pack writer does `fdatasync` + `DONTNEED` after
+  every completed layer. Explicit checkpoint prefetch is refused and a requested
+  multi-thread load is serialized. The final mmap-backed tensor per shard is cloned under the
+  same guard so the consumer cannot keep the whole file mapped. Pinned-arena allocations and
+  indivisible shard/layer I/O
+  keep a **16 GiB host MemAvailable floor** and **4 GiB hard cgroup `memory.max`
+  headroom** by default. `memory.high` remains an observable soft reclaim/throttle boundary;
+  it does not reduce the hard allocation budget. Tune the floors with
+  `VLLM_MOE_W2_MIN_MEM_AVAILABLE_GB` and
+  `VLLM_MOE_W2_MIN_CGROUP_HEADROOM_GB`; `VLLM_MOE_W2_CACHE_CONTROL=required` is the safe
+  default (`best-effort` or `off` are explicit unsafe overrides). Linux cgroup-v2 page cache is
+  accounted in `memory.current`; every check logs anon/file/file-mapped/swap/events plus its
+  available/headroom/transient budget for a cold-restage memory trace. In the all-43-layer
+  cold canary, cgroup memory.current peaked at 51.81 GB; observed anon/file maxima were
+  8.23/44.43 GB, host
+  `MemAvailable` stayed above 73.79 GB, cgroup swap and PSI/OOM counters remained zero,
+  and the host recorded no swap-out; host swap-in increased by 160 pages. The full canary used
+  the mechanism-equivalent `e7417054a6e8`
+  predecessor; the integrated `41d7b2f96ca3` image retains that P0 code and passed the exact-head
+  baked safety suites. See the [sanitized P0 receipts](evidence/public/ds4-w2-2026-07-11/p0/).
+- **Prefill is semantic, eager, and arena-safe.** The runner derives prefill from prompt
+  progress and propagates it explicitly through normal, replay, pipeline, gate, and DBO paths;
+  capture is refused for a real W2 prefill. Prefill disables draft-affinity/gate promotion,
+  holds exact per-layer BASE pins through the forward, and uses fill-only arena scan discipline.
+  Long prefills use the bulk base-only path; short prefills retain FP4 recovery without being
+  misclassified as decode. The arena's hot set persists to
   `<pack>.heat.json` and **preheats on boot** (57 GiB ≈ 35 s; `VLLM_MOE_W2_TIER_PREHEAT=0`
   opts out). Reader pool: `VLLM_MOE_W2_STORE_THREADS` (default 8).
+- **Tier managers run only between forwards.** A target step acquires the same exclusion lock
+  used by background promotion/eviction passes and holds it through target execution, BASE
+  replay, confidence-gate re-forward, and the pipeline barrier. Only the worker's final
+  `finish_forward_step()` releases the lock and wakes one manager pass. LRU timestamps use a
+  floating-point comparison key, and force-promote plus manager ticks consume immutable
+  layer-local seen sets, so saturated-LRU eviction and background tiering cannot mutate a live
+  route or overflow an integer timestamp.
 - **Observability:** with `VLLM_MOE_W2_DELTA_TRACE=1` the summary carries a
   `[base] tiered store: arena N/M | fetch rows X ram + Y nvme (Z% ram) | … p50/p99` line —
   the ram‑hit % *is* the arena‑coverage curve; grow `BASE_RAM_GB` if it sags.
-- Replays stay **bit‑identical** across backends (bytes are bytes; only the copy source
-  changes) — unit‑tested per backend × cold/warm/reboot/evict/overflow/scan/preheat in
+- Expert-store rows stay **byte-identical** across backends (only the copy source changes) —
+  unit-tested per backend × cold/warm/reboot/evict/overflow/scan/preheat in
   `tools/test_store_backends.py`.
 
 ---
@@ -424,8 +477,8 @@ Release **`baseline-2026-07-10`** — one row per supported recipe (`bench/recip
 <!-- bench:table:end -->
 
 ## Repository layout
-- **`patch/vllm-moet-v0.24.0.patch`** — the delta vs official vLLM `v0.24.0` (37 files,
-  +7.4k lines; applies clean on the tag). Goes with the pins above.
+- **`patch/vllm-moet-v0.24.0.patch`** — the single canonical delta vs official vLLM `v0.24.0`
+  (70 files, +14,018/-226 source lines; applies clean on the tag). Goes with the pins above.
 - **`Dockerfile.sm120-v024`** — the image: official `vllm/vllm-openai:v0.24.0` + patch + pins +
   cubins.
 - **`kernels/`** — SASS (`sass/`) + prebuilt SM120 cubins (`cubins-sm120/`, incl. the K=6144
diff --git a/docs/benchmarks/ds4-w2-5090-2026-07-11.md b/docs/benchmarks/ds4-w2-5090-2026-07-11.md
new file mode 100644
index 0000000..3384d92
--- /dev/null
+++ b/docs/benchmarks/ds4-w2-5090-2026-07-11.md
@@ -0,0 +1,354 @@
+# DeepSeek-V4-Flash W2 on one RTX 5090 — 2026-07-11
+
+The W2 lane now passes all three release priorities on one 32 GB RTX 5090:
+
+- **P0 passes:** the exact frozen release candidate completed an actual 46-shard, 43-expert-layer
+  cold restage under bounded memory. During the cold-restage window it retained a
+  103,726,841,856-byte host-available floor, used no cgroup or host swap, recorded no memory-limit
+  or OOM event, and restored the pre-existing serving surface with both protected pack
+  fingerprints unchanged.
+- **P1 passes at 32K:** the historical predecessor no-MTP configuration produced **119/120 exact-clean
+  answers and 0/120 sink detections under the frozen rules** across seeds 42, 43, and 44. All
+  **30/30 fixed prewarm prompts** passed without retry, and every seed passed both its pre-score
+  and post-score pool gate.
+- **P2 passes at 128K:** the comprehensive `41d7b2f` artifact produced **120/120 machine-exact and
+  semantically correct answers with 0/120 sink detections**, then retrieved exact needles at
+  120,000 prompt tokens at depths 0.1, 0.5, and 0.9. The frozen RC supplement then produced
+  **79/80 machine-exact, 80/80 semantically correct, and 0/80 sinks** and repeated the deepest
+  exact 120,000-token case.
+
+This is an empirical reliability result for the recorded sampler and pool policy. It is **not**
+a claim of bit-deterministic output, miss-free replay, or zero second-order residue.
+
+## Artifact lineage
+
+The evidence intentionally preserves four artifact roles rather than implying that every
+measurement came from one image:
+
+| Role | Source patch / image |
+|---|---|
+| P0 all-layer cold canary | mechanism-equivalent predecessor `e7417054a6e8`; measured memory belongs only to this canary |
+| Historical P1 32K receipts | patch `55d30bb9cf9bef45e7130fd5afe0090c5b540be22f6de3abfc51b782f738a6f7`; image `sha256:7385d21d26b665884e97a97dc67a100db328ed7b00b634e4d18f8aedd9f29eab` |
+| Comprehensive integrated P2 | official vLLM `ee0da84ab9e04ac7610e28580af62c365e898389`; patch `41d7b2f96ca3b966cac1b7ed5cff37bc03c27c616ed254aca961cc75a9ffe31d`; image `sha256:66abc2f145244e03ff0f0fcca088be813ad311f2963c62b281e4bb888ac605e9`; wrapper `5e69228e1e8e6f7345ae96657ab33b50e5805be7` |
+| Frozen RC, exact-head verification, and focused acceptance | origin/main cutoff `94be3aa3d7a8b82c7fc9687990a7edb6035f69f3`; official vLLM `ee0da84ab9e04ac7610e28580af62c365e898389`; patch `241ba984b1c56f5dc7adbc8d7f519d60b5746024bf7dfeb875e3546a668e79a7`; image `sha256:fc6e1244d60855fe45ccc0236daaaa722abcb8d354200eb92aca104bd954d3f2`; wrapper `0a73c2bcf262ba0aa53560d6290ad2af350e34fc` |
+
+The comprehensive `41d7b2f` patch changed 70 files with 14,018 insertions and 226 deletions; the
+frozen RC changes 70 files with 14,223 insertions and 226 deletions. Apply-check and
+whitespace-check pass, all 70 baked source files match the freshly applied tree, and the exact RC
+image passes the CUDA padded-route, store-safety, MXFP4 stream-safety, GPU store-backend, and
+base/full-FP4/split-FP4 numerical forward suites. Split-FP4 over the base cache is now supported
+behind its explicit flag with order-independent tier/base residency coupling. The default serving
+sentinel kept split mode off and used full-nibble FP4 planes; a separate exact-image E=8
+three-tier GPU fixture validated split mode and transient-miss/eviction behavior. The sanitized
+receipt is
+[`exact-image-verification.json`](../../evidence/public/ds4-w2-2026-07-11/exact-image-verification.json).
+
+## P0 — bounded cold restage
+
+**Outcome: the cold-restage wedge mechanism is bounded, and the exact frozen RC repeated the
+full cold-restage path.** The earlier mechanism canary completed all 43 routed-expert layers in
+249.2 seconds. Its historical measurements remain below for provenance.
+
+| Observed metric | Historical all-layer cold canary |
+|---|---:|
+| Samples | 245 |
+| Host `MemAvailable` minimum | 73,793,277,952 bytes |
+| Cgroup `memory.current` maximum | 51,807,191,040 bytes |
+| Cgroup memory peak | 51,807,522,816 bytes |
+| Anonymous-memory maximum | 8,231,129,088 bytes |
+| File-cache maximum | 44,430,127,104 bytes |
+| Cgroup swap maximum | 0 bytes |
+| Pressure-full `avg10` maximum | 0.0 |
+| `memory.high`, `memory.max`, OOM, and OOM-kill event deltas | all 0 |
+| Host swap-in / swap-out page deltas | 160 / 0 |
+| Host OOM-kill delta | 0 |
+| Production pack fingerprint after cleanup | unchanged |
+| Pre-existing serving surface after cleanup | restored and healthy |
+
+The frozen RC then cold-loaded all 46 checkpoint shards and built all 43 expert-pack layers. Shard
+loading completed in 203 seconds and the model became ready 340 seconds after container start.
+The cold-restage-only trace covers the first 220.7 seconds through completion of all checkpoint
+shards:
+
+| Observed metric | Frozen RC cold restage |
+|---|---:|
+| Samples | 218 |
+| Host `MemAvailable` minimum | 103,726,841,856 bytes |
+| Cgroup `memory.current` / peak maximum | 22,386,843,648 / 22,386,843,648 bytes |
+| Anonymous-memory maximum | 12,597,563,392 bytes |
+| File-cache maximum | 11,680,821,248 bytes |
+| Cgroup swap maximum | 0 bytes |
+| Pressure-full `avg10` maximum | 0.54 |
+| `memory.high`, `memory.max`, OOM, and OOM-kill event deltas | all 0 |
+| Host swap-in / swap-out page deltas | 0 / 0 |
+| Protected pack fingerprints after cleanup | both unchanged |
+| Pre-existing serving surfaces after cleanup | restored and healthy |
+
+The implementation fixes the original page-cache and staging failure mode at its source:
+
+1. MXFP4 expert tensors are materialized, quantized, durably written, and released one layer at
+   a time instead of accumulating the whole checkpoint.
+2. Released checkpoint shards and completed pack layers receive cache-drop advice; checkpoint
+   eviction is retried after the model consumer releases its final mapping.
+3. Explicit checkpoint prefetch is refused, requested parallel shard loading is serialized, and
+   an exact completed pack skips restaging.
+4. Each indivisible shard, layer, and arena operation preflights both host-available memory and
+   hard cgroup headroom. Required cache control fails closed; `memory.high` remains an observable
+   soft boundary rather than being mistaken for hard allocation capacity.
+5. The guarded cleanup path verifies that the temporary lane is gone, the original pack is
+   byte-identical by fingerprint, and the serving surface is healthy again.
+
+Evidence:
+
+- [Frozen RC cold-restage summary and trace](../../evidence/public/ds4-w2-2026-07-11/p0/frozen-rc-summary.json)
+  ([memory trace](../../evidence/public/ds4-w2-2026-07-11/p0/frozen-rc-memory-trace.tsv))
+- [All-layer cold-restage summary and trace](../../evidence/public/ds4-w2-2026-07-11/p0/cold-restage-summary.json)
+  ([memory trace](../../evidence/public/ds4-w2-2026-07-11/p0/cold-restage-memory-trace.tsv))
+- [Cleanup receipt](../../evidence/public/ds4-w2-2026-07-11/p0/cold-restage-cleanup-receipt.json)
+- [Constrained pack-build probe](../../evidence/public/ds4-w2-2026-07-11/p0/baked-constrained-pack-build.json)
+- [Checkpoint-residency and cache-drop probe](../../evidence/public/ds4-w2-2026-07-11/p0/baked-checkpoint-residency.json)
+- [Next-shard cache-drop retry probe](../../evidence/public/ds4-w2-2026-07-11/p0/baked-checkpoint-retry.json)
+
+### P0 evidence boundary
+
+The 245-sample table belongs only to mechanism-equivalent predecessor patch `e7417054a6e8`. The
+218-sample table belongs only to the exact frozen RC and ends after checkpoint-shard completion;
+it intentionally excludes the later quality and context workload. The final RC nevertheless ran
+the actual full 46-shard, 43-layer cold-restage lifecycle, passed the exact-image mechanism probes,
+and completed guarded cleanup. The two traces are retained separately rather than combining their
+maxima.
+
+## P1 — stable 32K quality and liveness
+
+**Outcome: P1 moves from 16/120 historical sinks to 0/120 sink detections under the frozen rules on the historical predecessor
+configuration, with 119/120 exact-clean answers.** Each seed independently meets the requested
+0–2 sink bar.
+
+The final configuration is a 32,768-token window, 8 GiB BASE cache, 6 GiB FP4 delta cache, LRU
+delta policy, confidence-gate tau 0.75, FP8 KV cache, MTP disabled, and one maximum sequence. The
+40-item evaluation uses fixed interleaving, temperature 0.6, top-p 0.95, a 700-token ceiling, no
+retries, and eval seeds 42–44.
+
+| Eval seed | Historical exact-clean | Historical sinks | Final exact-clean | Final sink detections | Fixed prewarm | Pre/post pool gates |
+|---:|---:|---:|---:|---:|---:|---|
+| 42 | 27/40 | 5/40 | 39/40 | **0/40** | 10/10 | pass / pass |
+| 43 | 22/40 | 7/40 | 40/40 | **0/40** | 10/10 | pass / pass |
+| 44 | 27/40 | 4/40 | 40/40 | **0/40** | 10/10 | pass / pass |
+| **Combined** | **76/120** | **16/120 (13.3%)** | **119/120** | **0/120** | **30/30** | **6/6 pass** |
+
+The lone final exact mismatch is seed 42 item `r17`: it ends with a mathematically equivalent
+formatted `3/8`, but the frozen exact-normalization rule does not accept that rendering. It is
+reported as non-exact and is not a sink; it is not silently promoted into the exact-clean total.
+
+The baseline is context, not a causal control arm. It is a corrected re-score of historical warm
+runs at tau 0.67, without the new fixed-prewarm and pool-lifecycle comparability contract. The
+final series uses tau 0.75 and predecessor provenance. The before/after table therefore establishes
+the operational delta, but it does not attribute that delta to one isolated knob. Baseline details
+are in the [sanitized corrected re-score](../../evidence/public/ds4-w2-2026-07-11/baseline/README.md)
+and [machine-readable aggregate](../../evidence/public/ds4-w2-2026-07-11/baseline/baseline-corrected.json).
+
+### Prewarm and pool gates
+
+Every final seed first passed the same ten-prompt `ds4-w2-prewarm-v4` suite at temperature 0,
+top-p 1, and fixed seed `20260711`, with no retry. The release policy was frozen before scoring at
+SHA-256 `1afb9b0d96d05c9d7a374dff88c94c76a3e854c8d046fb8ef1783275ad3f5630`.
+It requires full FP4 occupancy, zero unrestored BASE experts, a live FP4 tick/eviction count, and
+at least 16 additional evictions over each scoring run. All pre-gates and post-gates passed:
+
+| Eval seed | FP4 occupancy before/after | Unrestored BASE before/after | Post-score eviction delta | Post-score residue, latest 64-step window¹ |
+|---:|---:|---:|---:|---:|
+| 42 | 481/481 → 481/481 | 0 → 0 | 76,911 | 63/64 |
+| 43 | 481/481 → 481/481 | 0 → 0 | 80,902 | 64/64 |
+| 44 | 481/481 → 481/481 | 0 → 0 | 75,113 | 63/64 |
+
+¹ Residue is retained telemetry, not a release-policy rejection threshold; the approximation and
+its interpretation are explicit below.
+
+The frozen release policy is
+[`pool-gate-policy.json`](../../evidence/public/ds4-w2-2026-07-11/p1/pool-gate-policy.json).
+Per-seed manifests, raw rows, and prewarm receipts are published as
+[seed 42](../../evidence/public/ds4-w2-2026-07-11/p1/seed42.manifest.json),
+[seed 43](../../evidence/public/ds4-w2-2026-07-11/p1/seed43.manifest.json), and
+[seed 44](../../evidence/public/ds4-w2-2026-07-11/p1/seed44.manifest.json).
+
+### Explicit second-order approximation
+
+The BASE miss path first fetches every expert pair missed by the routed forward, then replays the
+step. Correcting early layers can change later routing and expose experts that the original pass
+never requested. At the default continuation threshold (`FP_THRESH=0`), those second-order experts are
+fetched for later steps, but their contribution to the current step remains zero. Consequently,
+`base_unrestored_experts = 0` proves fetch/store liveness; it does not prove a miss-free forward.
+
+A separately frozen strict diagnostic required zero residue. It passed 10/10 prewarm prompts but
+failed before scoring: 57 of the latest 64 steps carried second-order residue, despite 481/481 FP4
+occupancy and zero unrestored experts. The sanitized diagnostic is
+[`strict-zero-residue-diagnostic.json`](../../evidence/public/ds4-w2-2026-07-11/p1/strict-zero-residue-diagnostic.json).
+The run manifests therefore retain residue snapshots, while the final quality policy deliberately
+omits a residue rejection threshold. The three-seed frozen sink-detection result is the empirical
+acceptance test for that explicit approximation.
+
+This boundary matters: the final post-score windows still report second-order residue. The P1
+verdict is stable observed quality plus pool liveness, **not** bit-identical logits, identical token
+streams across boots, or mathematical equivalence to a fully resident forward.
+
+### Why the implementation stopped corrupting the run
+
+The predecessor patch fixed the implementation-level state hazards exposed by the earlier runs;
+the integrated patch retains them and adds the manager/forward exclusion described below:
+
+- Prefill is derived from prompt progress and propagated explicitly through normal, replay,
+  pipeline, gate, and microbatch paths. Real W2 prefill is eager; long prefills use the aligned
+  bulk path, while short prefills keep FP4 recovery instead of being misclassified as decode.
+- Padded CUDA-graph routes carry an explicit token-to-slot mapping, so padding cannot enter miss,
+  promotion, or output accounting. Routed rows are padded to the kernel block size and invalid
+  alignment fails loudly.
+- BASE residency uses an immutable, layer-local seen set. Prior main-stream work is drained before
+  reuse, only the previous layer's pins are cleared, and each current layer's hits/promotions stay
+  pinned through its forward. A later layer can no longer overwrite a shared snapshot or evict the
+  rows still in use.
+- Background tier managers use the same exclusion lock as the foreground forward. The target step
+  acquires it before routing, and the worker holds it through BASE replay, confidence-gate
+  re-forward, and the pipeline barrier; `finish_forward_step()` alone releases it and wakes one
+  manager pass. `step_begin()` only clears pins. LRU comparison keys are floating-point before
+  assigning infinity, and force-promote plus manager ticks receive immutable seen sets.
+- Draft-affinity prefetch and confidence-gate promotion are disabled during prefill, and real
+  prefill under capture is refused instead of silently taking decode semantics.
+
+### Near-edge context and lifecycle
+
+The same 32,768-token no-MTP server accepted **3/3 exact needles at 30,000 prompt tokens** at
+depths 0.1, 0.5, and 0.9. Tokenizer calibration and response usage both had to report exactly
+30,000 tokens, and each terminal passphrase had to match exactly. This is P1 near-edge validation
+and a rehearsal of the P2 harness; it is not a 128K result. See the
+[context manifest](../../evidence/public/ds4-w2-2026-07-11/p1/context-30k.manifest.json) and
+[receipts](../../evidence/public/ds4-w2-2026-07-11/p1/context-30k.receipts.jsonl).
+
+Across scoring and context, cgroup swap, hard-limit, and OOM events stayed at zero. The 96 GiB
+soft `memory.high` boundary was crossed narrowly for 13 samples and generated reclaim events;
+PSI full `avg10` peaked at 0.49. The guarded lane retained at least 4,294,172,672 bytes of hard
+headroom, then stopped cleanly, preserved both pack fingerprints, restored both services, and
+removed the 230,317,622,095-byte disposable store. Public receipts:
+[aggregate](../../evidence/public/ds4-w2-2026-07-11/p1/aggregate.json),
+[runtime and memory audit](../../evidence/public/ds4-w2-2026-07-11/p1/runtime-clean.json), and
+[cleanup](../../evidence/public/ds4-w2-2026-07-11/p1/cleanup.json).
+
+### Excluded MTP diagnostic
+
+The MTP-on investigation crossed noncanonical patch/image identities and used a different runtime
+shape, so it contributes **zero rows** to the final 120-response series. Its private diagnostic
+receipts are not used to support any public result. Combining it with the historical no-MTP seeds
+would erase both the provenance and gate boundary; this report does not do so.
+
+### Excluded first integrated candidate
+
+The first upstream-integrated candidate (`ec4640f` patch generation, `5467a7ff` image prefix)
+reached a healthy 128K server but failed frozen prewarm at item 3 before scoring. It contributes
+zero quality rows and no context receipt. That failure exposed a background tier-manager pass
+overlapping the live target/replay path. Intermediate patch `4708c9d41b50` added the shared
+forward-exclusion lifecycle and passed the full gate. The comprehensive `41d7b2f96ca3` union then
+integrated the newer upstream split-FP4 work, preserved that exclusion and restage safety, made
+split mode fail closed over the base cache at that generation, and reran the entire gate from a
+fresh namespace. The later frozen RC adds explicit split/base residency coupling and validates it
+separately while leaving split mode off in the serving acceptance run.
+
+## P2 — 128K context
+
+**Outcome: the comprehensive `41d7b2f` single-5090 configuration serves a validated 131,072-token window and
+retrieves 3/3 exact needles at 120,000 prompt tokens without weakening the quality gate.** This is
+the largest context tested in this lane. No decode- or prefill-throughput claim is attached to it.
+
+The P2 runtime keeps the P1 BASE/FP4 policy—8 GiB BASE, 6 GiB delta, LRU, tau 0.75, FP8 KV—but
+sets `--max-model-len 131072`, disables MTP, and permits one maximum sequence. Model loading used
+22.87 GiB of device memory. The engine exposed 720,677 KV tokens and reported 5.50× capacity for
+a full 131,072-token request. The exact image reached health 200, completed the controller's
+30-second hold, and
+entered scoring with zero start-to-ready swap, memory-high, hard-limit, or OOM events. See the
+[readiness receipt](../../evidence/public/ds4-w2-2026-07-11/p2/readiness.json) and
+[sanitized server identity](../../evidence/public/ds4-w2-2026-07-11/p2/server.json).
+
+### 128K quality
+
+Each seed repeated the same frozen 40-item sampler and ten-prompt prewarm used for P1:
+
+| Eval seed | Machine exact | Semantic equivalent | Sink detections | Fixed prewarm | Pre/post gates | Score-time evictions |
+|---:|---:|---:|---:|---:|---|---:|
+| 42 | 40/40 | 40/40 | **0/40** | 10/10 | pass / pass | 42,760 |
+| 43 | 40/40 | 40/40 | **0/40** | 10/10 | pass / pass | 43,118 |
+| 44 | 40/40 | 40/40 | **0/40** | 10/10 | pass / pass | 44,469 |
+| **Combined** | **120/120** | **120/120** | **0/120** | **30/30** | **6/6 pass** | **130,347** |
+
+Seed 44 item `c18` returned `(2,5)` for expected `(2, 5)`. The frozen exact scorer normalizes that
+tuple whitespace and counts it exact; the legacy `lenient` diagnostic does not. It is not a sink,
+and the machine total remains 120/120. All 120 responses were HTTP 200 with
+`finish_reason=stop`, and no retries or request errors occurred. The public
+[quality aggregate](../../evidence/public/ds4-w2-2026-07-11/p2/quality-aggregate.json) and
+[per-seed manifests for seed 42](../../evidence/public/ds4-w2-2026-07-11/p2/seed42.manifest.json),
+[43](../../evidence/public/ds4-w2-2026-07-11/p2/seed43.manifest.json), and
+[44](../../evidence/public/ds4-w2-2026-07-11/p2/seed44.manifest.json) retain the machine/semantic
+boundary and pool snapshots.
+
+### Exact 120K retrieval
+
+| Needle depth | Tokenizer count | Response usage | Exact answer | Request wall time |
+|---:|---:|---:|---|---:|
+| 0.1 | 120,000 | 120,000 | pass | 250.041 s |
+| 0.5 | 120,000 | 120,000 | pass | 248.028 s |
+| 0.9 | 120,000 | 120,000 | pass | 245.272 s |
+
+All three admitted receipts were fresh zero-tolerance requests. Each case calibrated to exactly
+120,000 tokens through `/tokenize` before its sole inference; no rejected calibration contributed
+a chat response in this run. Response usage also read 120,000 and each terminal passphrase
+matched exactly. The public
+[composite manifest](../../evidence/public/ds4-w2-2026-07-11/p2/context-120k.manifest.json) and
+[selected receipts](../../evidence/public/ds4-w2-2026-07-11/p2/context-120k.receipts.jsonl)
+document the selection boundary.
+
+Retrieval success is not a miss-free-forward claim. The three requests emitted interim KPI
+windows with 42, 15, and 13 residue steps respectively, then each closed with a zero-replay,
+zero-residue latest window. The exact answers empirically validate retrieval under the declared
+approximation; they do not make every internal forward miss-free.
+
+### Frozen RC acceptance supplement
+
+The exact frozen RC reran the two quality seeds and deepest context case most sensitive to the
+final split/base-coupling and padded-route hardening delta. It did not relabel or duplicate the
+comprehensive `41d7b2f` seed-44 or depth-0.1/0.5 receipts.
+
+| Eval seed | Machine exact | Semantic equivalent | Sink detections | Fixed prewarm | Pre/post gates | Score-time evictions |
+|---:|---:|---:|---:|---:|---|---:|
+| 42 | 40/40 | 40/40 | **0/40** | 10/10 | pass / pass | 46,560 |
+| 43 | 39/40 | 40/40 | **0/40** | 10/10 | pass / pass | 44,345 |
+| **Combined** | **79/80** | **80/80** | **0/80** | **20/20** | **4/4 pass** | **90,905** |
+
+All 80 responses were HTTP 200 with `finish_reason=stop`, with no prewarm retry and no unrestored
+BASE expert. Seed 43 item `r17` returned exact mathematical content as LaTeX
+`\(\frac{3}{8}\)` for expected `3/8`; the strict machine scorer correctly kept it out of the
+machine-exact total while the semantic scorer accepted it. It was not a sink.
+
+The RC then calibrated and issued one depth-0.9 request whose tokenizer count and response usage
+were both exactly 120,000 tokens. It returned the exact needle with HTTP 200 and
+`finish_reason=stop` in 250.673 seconds. Two earlier zero-tolerance candidates were rejected by
+tokenizer-only calibration, so neither produced a chat request.
+
+Across the complete 1,744.3-second RC lifecycle, cgroup and host swap stayed at zero, no
+hard-limit or OOM event occurred, and hard cgroup headroom stayed at or above 4,294,537,216 bytes.
+The 96 GiB soft boundary was crossed narrowly and generated 19,598 `memory.high` events; that
+reclaim is reported rather than called zero. Cleanup removed the disposable payload, preserved
+both protected pack fingerprints, and restored both serving surfaces at HTTP 200. See the
+[frozen RC sentinel](../../evidence/public/ds4-w2-2026-07-11/p2/frozen-rc-sentinel.json).
+
+### Comprehensive `41d7b2f` cleanup
+
+After the final exact receipt, the guarded controller stopped the isolated container, found both
+production and immutable-seed pack fingerprints byte-identical, restored both serving services at
+HTTP 200, and removed only the 230,317,622,135-byte disposable pack copy. The 40-byte increase
+over the predecessor is the two `fp4_split=false` metadata fields; both payload pack sizes are
+unchanged. Cgroup swap, hard-limit, OOM, and OOM-kill events stayed at zero. Host `MemAvailable`
+never fell below 103,752,298,496 bytes, and hard cgroup headroom never fell below 4,294,488,064
+bytes. File-cache growth crossed the 96 GiB soft `memory.high` boundary for two sampled instants,
+produced 28,089 soft-high events, and recorded PSI full `avg10` up to 0.79. That
+reclaim/pressure is reported explicitly rather than
+called zero. Immediate post-restart curls briefly saw connection refusal; the controller completed
+cleanup and later follow-up checks returned HTTP 200/200. See the
+[runtime audit](../../evidence/public/ds4-w2-2026-07-11/p2/runtime-clean.json) and
+[cleanup receipt](../../evidence/public/ds4-w2-2026-07-11/p2/cleanup.json).
diff --git a/docs/quality.md b/docs/quality.md
index 5da8fac..2330088 100644
--- a/docs/quality.md
+++ b/docs/quality.md
@@ -18,6 +18,71 @@ runs on the *same* serving stack (only `MODEL_DIR` differs).
 3. **Arithmetic** — 5 multi‑step problems (the official model itself tops out at 3/5).
 4. **Long context** — passphrase needle at 100K/250K/440K depths (`tools/needle_probe.py`).
 
+## DS4-W2 reliability methodology (P1/P2 release gate)
+
+The release gate uses `tools/ds4_eval/`; it is stricter than a transcript spot check. Every warm
+run must first pass the fixed ten-prompt `ds4-w2-prewarm-v4` suite (temperature 0, top-p 1, seed
+`20260711`, independent of the eval seed). There are no hidden retries: the first bad response
+aborts and its receipt remains on disk. The run manifest binds the result to immutable server
+provenance, including the boot, container and image identities, complete source-diff hash,
+checkpoint and pack fingerprints, launcher hash, runtime argv, W2 environment, and engine settings.
+
+Before looking at quality results, register one pool policy and retain its SHA-256. The same frozen
+policy evaluates the configured live metrics before and after scoring. The release policy requires
+full FP4 occupancy, zero unrestored BASE experts, and measured eviction progress so a stale
+saturated pool cannot pass; replay and second-order residue remain recorded in every snapshot even
+when they are not rejection thresholds. A failed pre- or post-gate invalidates the run regardless
+of its answers. At the default continuation threshold (`FP_THRESH=0`), second-order experts are
+fetched for later steps but their current-step contributions remain zero, so `UNRESTORED=0` is a
+mechanism-liveness gate rather than a miss-free or bit-deterministic claim. The three-seed quality
+receipts are the empirical acceptance test for that explicit throughput approximation.
+
+P1 scores the fixed 40-item set separately at eval seeds 42, 43, and 44 (120 responses total).
+Exact clean correctness is primary: the normalized terminal answer must match and the response
+must not be a sink. The answer-anywhere `lenient` score is diagnostic only and never rescues a
+sink. Sink rules catch max-token non-completions, repeated 3/4-grams, repeated lines, collapsed
+vocabulary, and special-token spew; duplicate IDs, missing rows, or conflicting token-count aliases
+fail closed. The corrected historical tau-0.67 baseline is **16/120 sinks (13.3%)** across those
+three seeds. It is context only, not a tau-0.75 control: its warmups and pool evidence do not meet
+the new comparability contract.
+The corrected baseline summary is published under
+[`evidence/public/ds4-w2-2026-07-11/baseline/`](../evidence/public/ds4-w2-2026-07-11/baseline/).
+
+Use repository-relative inputs and a localhost endpoint; write each seed to a new output directory:
+
+```bash
+PORT="${PORT:-18001}"
+python3 tools/ds4_eval/eval_rig.py \
+  --items tools/ds4_eval/items.json \
+  --server-provenance evidence/example-p1/server.json \
+  --pool-gate-policy evidence/example-p1/pool-gate.json \
+  --pool-command-json '["docker","logs","ds4-w2-candidate"]' \
+  --output-dir evidence/example-p1/seed-42 \
+  --run-label example-p1-s42 \
+  --url "http://localhost:${PORT}/v1/chat/completions" \
+  --model deepseek-v4-flash-w2 --mode warm --eval-seed 42 \
+  --eval-temperature 0.6 --eval-top-p 0.95 --eval-max-tokens 700 \
+  --expected-count 40
+```
+
+Repeat with seeds 43 and 44 and distinct labels/directories. P2 is a two-part contract on the same
+ready server: first run this quality gate, then run `tools/ds4_eval/context_probe.py` with MTP off,
+`--max-num-seqs 1`, a validated 131,072-token window, and 120,000-token needles at depths 0.1,
+0.5, and 0.9. The probe requires exact tokenizer/usage token counts and an exact terminal
+passphrase. A successful 128K boot without retrieval, or retrieval without the quality gate, is
+not a P2 verdict.
+
+The guarded single-RTX-5090 application of this protocol is recorded in
+[`ds4-w2-5090-2026-07-11.md`](benchmarks/ds4-w2-5090-2026-07-11.md), with sanitized machine
+receipts under [`evidence/public/ds4-w2-2026-07-11/`](../evidence/public/ds4-w2-2026-07-11/).
+The historical predecessor P1 series at 32K produced 119/120 machine-exact and 120/120
+semantically correct answers with 0/120 frozen-rule sink detections. The current integrated
+artifact's independent P2 series at 128K improved to 120/120 machine-exact and semantically
+correct, with 0/120 sink detections, thereby satisfying the P1 stability gate at the larger
+window. P2 then passed exact 120,000-token needles at depths 0.1, 0.5, and 0.9. Each case was
+calibrated with tokenizer-only requests before its sole inference; the public composite contains
+only fresh zero-tolerance receipts whose tokenizer and response-usage counts both equal 120,000.
+
 ## Bits‑vs‑quality ablation (same stack, only the expert codes change)
 | codebook | bits | MTP acc. length | draft accept % | arithmetic | coherence |
 |---|---|---|---|---|---|
diff --git a/docs/v024-port.md b/docs/v024-port.md
index b2079f7..b8b08e9 100644
--- a/docs/v024-port.md
+++ b/docs/v024-port.md
@@ -2,7 +2,7 @@
 
 The project targets **official vLLM v0.24.0**, which ships DeepSeek‑V4 + SM120 natively
 (`vllm/models/deepseek_v4/`, FlashInfer SM120 sparse‑MLA, GLM‑5.x `GlmMoeDsaForCausalLM`).
-Our overlay is a **7.4k‑line patch** (37 files): the 2‑bit expert planes, the FP4 delta
+Our overlay is a **14.0k-line patch** (70 files, 14,018 insertions): the 2-bit expert planes, the FP4 delta
 cache, the confidence gate, the cubit dispatch, the expert stores — plus the SM120 fixes
 below.
 
@@ -40,9 +40,10 @@ Environment pins that go with the patch (both required on SM120):
 
 ## Our hooks
 
-- `mxfp4.py` (`Mxfp4MoEMethod`) — FP4‑checkpoint path (DeepSeek‑V4‑Flash): host‑stage experts
-  at `create_weights`, build 2‑bit planes at `process_weights_after_loading`, `moe_w2_forward`
-  in `apply`.
+- `mxfp4.py` (`Mxfp4MoEMethod`) — FP4-checkpoint path (DeepSeek-V4-Flash): create zero-sized
+  expert parameters, materialize one guarded layer while its tensors load, build/write its
+  2-bit planes once, then release it before the next layer. Exact pack identity skips the
+  materialization; `moe_w2_forward` runs in `apply`.
 - `fp8.py` (`Fp8MoEMethod`) — FP8 block‑quant checkpoint path (DS4‑Flash‑Base,
   **GLM‑5.2‑FP8**): same three hooks; the loader re‑quantizes fp8+f32‑block‑128 to the
   sign‑symmetric 2‑bit codebook at load (`build_layer_planes_fp8`, float64 math, golden‑tested
@@ -61,8 +62,9 @@ Environment pins that go with the patch (both required on SM120):
   RAM; the GPU pool caches hot experts through the same slot‑table/manager/eviction machinery
   as the delta tier. Decode misses zero the pair's contribution, bump an in‑graph miss
   counter, and the runner fetches all missing routed experts synchronously (batched pinned
-  H2D, 51.6 GiB/s measured) and replays the step's graph once — replay bit‑identical to a
-  resident forward (unit‑tested, `internal` test_base_cache). Prefill prefetches per layer
+  H2D, 51.6 GiB/s measured) and replays the step's graph once. Replay is bit-identical to a
+  resident forward only when the replay creates no second-order misses (the zero-residue case
+  is unit-tested). Prefill prefetches per layer
   via `ensure_resident`. TP MAX‑reduces the miss decision. Under **PP** a miss is local to
   its stage (per‑stage counter, inputs still held in the stage's static buffers), so each
   stage re‑runs only its own **segment** before activations flow downstream — no cross‑stage
@@ -75,7 +77,10 @@ Environment pins that go with the patch (both required on SM120):
   `VLLM_MOE_W2_FP_MAX` bounds ping‑pong). An unconditional loop collapsed decode at low
   coverage (GLM TP2 29→16 tok/s, DS4‑14 GiB 43→15 — chasing a moving target at up to 8
   forwards/step); the live A/B behind the default: thresh 16 → 19.2, thresh 0 → 28.3 tok/s
-  at identical quality probes. Slots touched by any pass of a step are pinned against
+  at identical quality probes. With the default threshold, second-order experts are fetched
+  for later steps but their current-step contributions remain zero; `fp-residue` reports this
+  explicit approximation, so zero `UNRESTORED` is not a miss-free or determinism claim.
+  Slots touched by any pass of a step are pinned against
   eviction until the next step (the passes must not cannibalize each other's fetches); on
   tight pools an emergency eviction pass (synchronous callers only) relaxes the 2‑tick
   coldness bound rather than leave a miss UNRESTORED. Coexists with the FP4 need‑pool (explicit
@@ -94,9 +99,10 @@ Environment pins that go with the patch (both required on SM120):
   | 11 GiB (util 0.90) | 15.2% | 96.5–97.7% | 32.7 / **27–28 tok/s** |
   | 14 GiB (util 0.95) | 19.3% | 98.7–98.9% | 43.4 / **~31 tok/s** |
 
-  The pre‑fix column is the single‑replay stack (silently kept second‑order zeros — the
-  nondeterminism the fixed‑point restore later closed); the shipped column is the final
-  adaptive‑replay + NVMe‑store stack (2026‑07‑10 evening, same box/bench idiom). The pool
+  The pre-fix column is the single-replay stack (silently kept second-order zeros); the
+  shipped column is the adaptive-replay + NVMe-store stack, which measures that residue and
+  can chase it when `FP_THRESH` is raised but accepts it at the throughput-oriented default
+  (2026-07-10 evening, same box/bench idiom). The pool
   slope survives the stack change (+12–15% for 3 GiB). 14 GiB does NOT fit at util 0.90 (KV
   needs 0.46 GiB after graphs) — raise `--gpu-memory-utilization` alongside the pool. The
   engine logs pool
@@ -120,23 +126,60 @@ Environment pins that go with the patch (both required on SM120):
   pinned 33.0 / pack 25.5 / tiered+20 GiB arena **32.8 tok/s (parity)** at RSS 26–33 vs
   42–44 GiB; GLM TP2 both stores on NVMe: expert‑store RAM ~568 → ~136 GiB, 28–32 tok/s
   (tol 0–8, adaptive replay), needle 4/4 to **121K prompt tokens** at a served 128K window
-  (KV 157K tokens measured). Prefill batches are scan‑flagged (fill free arena slots,
-  never evict the decode hot set — a caller flag, NOT a batch‑size heuristic: GLM's
-  100+‑row decode replay fetches misclassified and froze the arena at −66%); the arena hot
-  set persists to `<pack>.heat.json` and preheats on boot (57 GiB ≈ 35 s). **Boot‑from‑pack**
+  (KV 157K tokens measured). Prefill is derived authoritatively from each request's
+  `num_computed_tokens < num_prompt_tokens` state and carried as
+  `ForwardContext.has_prefill` through normal, replay, pipeline, gate, and DBO execution.
+  A real W2 prefill is eager (capture fails loud), disables draft-affinity prefetch and
+  confidence-gate promotion, and uses fill-only arena reads so it cannot evict the decode
+  hot set. Long prefills use the aligned bulk base-only path; short prefills retain FP4
+  recovery and mblock-4 instead of falling into the decode path. Padded CUDA-graph routes
+  carry an explicit token-slot mapping and are masked out of miss, promotion, and output
+  accounting. The arena hot set persists to `<pack>.heat.json` and preheats on boot
+  (57 GiB approximately 35 s). **Boot-from-pack**
   (`_try_skip_requant`, all three loaders): a layer present in every serving pack skips
   dequant→re‑quant entirely — GLM TP2 second boot 408 s vs ~11 min, no ~405 GiB transient;
   the pack is a persistent quantization cache keyed by shape/config sidecar match. Also
-  fixed here, exposed by long‑prefill testing but **pre‑existing**: the manager tick and
-  forward‑thread paths ran concurrent seen‑snapshots into one shared pinned `_seen_host`
-  (torch's two‑pass `nonzero` overruns its output when the input mutates mid‑call →
-  TensorAdvancedIndexing.cpp:3008 assert → glibc heap corruption → dead worker on ≥16K
-  prefills; a torn snapshot could also evict an in‑flight expert's slot). Snapshots are now
-  serialized under a dedicated lock. Backends unit‑tested byte‑identical
-  (`tools/test_store_backends.py`: 3 backends × cold/warm/reboot/evict/overflow/scan/
+  fixed here, exposed by long-prefill testing but **pre-existing**: manager and forward
+  threads previously reused one mutable pinned seen-mask, so a later layer could overwrite
+  the snapshot while the current layer selected or fetched rows. Each residency operation
+  now drains prior main-stream work, freezes an immutable layer-local `seen_set`, clears only
+  the previous BASE layer's pins, and pins the current layer's hits/promotions until its
+  forward completes. A target step also acquires the exclusion lock shared with each background
+  tier-manager pass. That lock stays held through target execution, BASE replay, confidence-gate
+  re-forward, and the pipeline barrier; only `finish_forward_step()` releases it and wakes one
+  manager pass. `step_begin()` therefore clears pins without racing a wake. Saturated LRU uses a
+  floating-point timestamp key before assigning infinity, and both force-promote and manager
+  ticks receive immutable seen sets. Backends unit-tested byte-identical
+  (`tools/test_store_backends.py`: 3 backends x cold/warm/reboot/evict/overflow/scan/
   preheat, both IO modes). Ops: packs on a bind‑mounted real FS (not overlayfs); ~1 TB NVMe
   for the full GLM stack; parity holds even on a Gen3‑x4 drive (3.7 GB/s) — steady‑state
   misses ride the page cache, cold shifts pay drive speed.
+  Cold rebuilds are bounded separately from steady reads: MXFP4 uses guarded one-layer
+  materialization/build/write/release rather than retaining every layer's expert tensors;
+  pack-skip and completion guards prevent unnecessary or duplicate work. Eviction is
+  requested for each
+  released safetensors shard and each durable pack layer, then retried after the model
+  consumer unwinds,
+  checkpoint prefetch is refused, requested multi-thread shard loading is serialized, and
+  arena/shard/layer operations preflight a default 16 GiB host + 4 GiB hard cgroup
+  `memory.max` reserve. `memory.high` is traced as a soft reclaim/throttle boundary and does
+  not reduce that hard allocation budget. The
+  safety controls are `VLLM_MOE_W2_CACHE_CONTROL`,
+  `VLLM_MOE_W2_MIN_MEM_AVAILABLE_GB`, and
+  `VLLM_MOE_W2_MIN_CGROUP_HEADROOM_GB`; cache control defaults to fail-closed `required`.
+  The all-43-layer cold canary completed with a 51,807,191,040-byte cgroup peak,
+  8,231,129,088-byte anon peak, 44,430,127,104-byte file peak, and a
+  73,793,277,952-byte host `MemAvailable` floor; cgroup swap, PSI, limit, and OOM events were
+  zero, and the host recorded no swap-out; host swap-in increased by 160 pages.
+  That all-layer canary used predecessor patch `e7417054a6e8`; integrated patch
+  `41d7b2f96ca3` retains the mechanism and passed the exact-head baked safety tests. Sanitized
+  trace and cleanup receipts are under
+  [`evidence/public/ds4-w2-2026-07-11/p0/`](../evidence/public/ds4-w2-2026-07-11/p0/).
+  Historical predecessor receipts establish the 32K P1 lane; the current integrated image's
+  guarded 128K series independently satisfies the same stability gate. Both are documented in
+  [`ds4-w2-5090-2026-07-11.md`](benchmarks/ds4-w2-5090-2026-07-11.md): 0/120 frozen-rule sink
+  detections in each series, with exact 120,000-token retrieval at depths 0.1, 0.5, and 0.9 on
+  the integrated image. This is a quality/context result; no 128K throughput is claimed.
 - **Deterministic unpermute**: the MoE output scatter used atomic `index_add_`, so identical
   runs wobbled (~1.6e‑2 on prefill) and greedy decode was not reproducible (surfaced by the
   PP determinism investigation; never PP‑specific). Valid `sorted_ids` form a permutation of
diff --git a/evidence/public/ds4-w2-2026-07-11/README.md b/evidence/public/ds4-w2-2026-07-11/README.md
new file mode 100644
index 0000000..d556ec7
--- /dev/null
+++ b/evidence/public/ds4-w2-2026-07-11/README.md
@@ -0,0 +1,100 @@
+# DS4 W2 guarded validation evidence — 2026-07-11
+
+This directory contains the sanitized, publishable receipts for the single-RTX-5090
+DeepSeek-V4-Flash W2 safety, quality, and context run. The measured narrative and evidence
+boundaries are in [`docs/benchmarks/ds4-w2-5090-2026-07-11.md`](../../../docs/benchmarks/ds4-w2-5090-2026-07-11.md).
+
+## Artifact lineage
+
+| Evidence role | Artifact identity |
+|---|---|
+| P0 all-layer cold canary | mechanism-equivalent predecessor `e7417054a6e8`; measured memory belongs only to this canary |
+| Historical P1 32K receipts | patch `55d30bb9cf9bef45e7130fd5afe0090c5b540be22f6de3abfc51b782f738a6f7`; image `sha256:7385d21d26b665884e97a97dc67a100db328ed7b00b634e4d18f8aedd9f29eab` |
+| Comprehensive integrated P2 predecessor | official vLLM `ee0da84ab9e04ac7610e28580af62c365e898389`; patch `41d7b2f96ca3b966cac1b7ed5cff37bc03c27c616ed254aca961cc75a9ffe31d`; image `sha256:66abc2f145244e03ff0f0fcca088be813ad311f2963c62b281e4bb888ac605e9`; wrapper `5e69228e1e8e6f7345ae96657ab33b50e5805be7` |
+| Frozen release candidate and focused current-head supplement | official vLLM `ee0da84ab9e04ac7610e28580af62c365e898389`; origin/main cutoff `94be3aa3d7a8b82c7fc9687990a7edb6035f69f3`; patch `241ba984b1c56f5dc7adbc8d7f519d60b5746024bf7dfeb875e3546a668e79a7`; image `sha256:fc6e1244d60855fe45ccc0236daaaa722abcb8d354200eb92aca104bd954d3f2`; wrapper `0a73c2bcf262ba0aa53560d6290ad2af350e34fc` |
+
+[`exact-image-verification.json`](exact-image-verification.json) records the clean-apply shape,
+baked source hashes, exact-image test matrix, and synthetic safety-probe hashes for the frozen
+release candidate. The earlier `41d7b2f` artifact remains the comprehensive three-seed,
+three-depth quality record; the frozen candidate adds a focused exact-head release sentinel.
+
+## P0 — bounded cold restage
+
+[`p0/`](p0/) contains the all-43-layer cold-restage memory trace and guarded cleanup receipt.
+That measured canary ran on mechanism-equivalent predecessor patch `e7417054a6e8`; its memory
+maxima are not represented as integrated-image measurements. The checkpoint-residency,
+next-shard retry, and constrained-pack probes in the same directory were rerun on the frozen
+release-candidate image and bind to its exact image ID.
+
+The canary kept cgroup swap, pressure, hard-limit, and OOM events at zero and recorded no host
+swap-out. Host swap-in increased by 160 pages; that boundary is explicit in the summary and trace.
+
+The frozen candidate also received a fresh full cold restage. [`p0/frozen-rc-summary.json`](p0/frozen-rc-summary.json)
+records 46/46 checkpoint shards, all 43 expert-pack layers, zero swap and OOM activity, zero
+memory-limit event deltas, preserved production and immutable-seed fingerprints, restored service
+health, and disposable-store removal. Its 218-sample cold-only trace is
+[`p0/frozen-rc-memory-trace.tsv`](p0/frozen-rc-memory-trace.tsv). These measurements bind to the
+frozen candidate and supersede the historical limitation only for that candidate; the earlier
+canary files remain historical receipts.
+
+## P1 — 32K quality and near-edge context
+
+The files under [`p1/`](p1/) are retained byte-for-byte as historical predecessor evidence.
+[`p1/aggregate.json`](p1/aggregate.json) reports 119/120 machine-exact, 120/120 semantically
+correct, and 0/120 frozen-rule sink detections across seeds 42–44, plus 30/30 prewarm prompts and
+all six pool-lifecycle gates. Per-seed manifests, raw responses, and warmup receipts are retained
+beside the aggregate. Published receipt hashes are recomputed after sanitization.
+
+[`p1/context-30k.manifest.json`](p1/context-30k.manifest.json) and
+[`p1/context-30k.receipts.jsonl`](p1/context-30k.receipts.jsonl) record 3/3 exact needles at
+30,000 prompt tokens. [`p1/runtime-clean.json`](p1/runtime-clean.json) records the complete lane
+through context, including soft `memory.high` reclaim and nonzero PSI; [`p1/cleanup.json`](p1/cleanup.json)
+records fingerprint preservation, service restoration, and disposable-store removal.
+
+## P2 — comprehensive 128K series and frozen-RC sentinel
+
+The comprehensive P2 receipts bind to the `41d7b2f` integrated predecessor above. Its three-seed
+128K series also satisfies the P1 stability requirement.
+[`p2/quality-aggregate.json`](p2/quality-aggregate.json)
+reports 120/120 machine-exact and semantically correct, with 0/120 frozen-rule sink detections at
+a 131,072-token window.
+
+[`p2/context-120k.manifest.json`](p2/context-120k.manifest.json) and
+[`p2/context-120k.receipts.jsonl`](p2/context-120k.receipts.jsonl) contain the admitted exact
+120,000-token receipts at depths 0.1, 0.5, and 0.9. All three published receipts use zero token
+tolerance. Each case calibrated to exactly 120,000 tokens with tokenizer-only requests before its
+sole inference; no rejected calibration contributed a chat response in this run.
+Readiness, complete runtime/memory, and guarded cleanup are recorded in
+[`p2/readiness.json`](p2/readiness.json), [`p2/runtime-clean.json`](p2/runtime-clean.json), and
+[`p2/cleanup.json`](p2/cleanup.json).
+
+[`p2/frozen-rc-sentinel.json`](p2/frozen-rc-sentinel.json) is the focused supplement for the
+frozen candidate. It records two quality seeds (80 responses): 79/80 strict machine-exact,
+80/80 semantically correct, 0/80 sink detections, all 20 prewarms without retry, and all four
+pool gates. The sole strict mismatch is an explicitly recorded equivalent LaTeX rendering of
+`3/8`. It also records one exact 120,000-token retrieval at depth 0.9 with zero token tolerance,
+plus complete-run memory and cleanup results. This supplement did not repeat seed 44 or depths
+0.1 and 0.5, so those broader claims remain attached only to the immediately preceding
+comprehensive P2 artifact.
+
+The frozen source supports split-FP4 refinement over a base cache only behind the explicit split
+flag, with coupled base/refinement residency. The serving sentinel kept split mode disabled and
+used full-nibble FP4 delta planes. Split mode was exercised separately by the exact-image E=8 GPU
+three-tier fixture; serving-sentinel results are not presented as live split-mode evidence.
+
+## Interpretation and sanitization
+
+These receipts establish bounded restaging, observed quality, pool liveness, exact selected
+retrieval, and guarded cleanup for the recorded artifact and policy. They do not establish
+bit-deterministic output, a miss-free forward, numerical equivalence to fully resident FP4, or
+128K throughput. Second-order residue remains explicit.
+
+Host topology, opaque runtime IDs, host-specific paths, host- and address-specific endpoint
+details, checkpoint/pack
+fingerprints, launcher identity, and response IDs remain in private source receipts. Public run
+labels are descriptive. The lane started on July 11 Pacific time, while UTC timestamps in the
+integrated receipts fall on July 12. Verify the final public tree from this directory with:
+
+```bash
+shasum -a 256 -c SHA256SUMS
+```
diff --git a/evidence/public/ds4-w2-2026-07-11/SHA256SUMS b/evidence/public/ds4-w2-2026-07-11/SHA256SUMS
new file mode 100644
index 0000000..0be7b77
--- /dev/null
+++ b/evidence/public/ds4-w2-2026-07-11/SHA256SUMS
@@ -0,0 +1,47 @@
+4e894b6b293a3f831cd5ce830d03f691e09a91d26b27b3a329d88fcf30f64c9f  ./README.md
+14e410651bc505b1bec6ec6a439be1d19600987a797fc6f615f71d8bb79e489a  ./baseline/README.md
+85b3eff601c3a8d1f70810fc297fa8f6a40c7e5e5b238710cc1068e86ffe7054  ./baseline/baseline-corrected.json
+be456388a14da84f0a2d7f5fd30222fddeac10d3847946f6f6fc259b2c12bbd4  ./exact-image-verification.json
+e7eb19fc75dc8b49d587b9f07c87c5db182cd764a2f85998aac78d85a03698b8  ./p0/baked-checkpoint-residency.json
+27de4d773ac034dce3d9f6d29b70275ec195219fd69077343a47843dc764c107  ./p0/baked-checkpoint-retry.json
+28d4b707ff1f49cf3b52091483a47d20ae855f001bd9937b48a69dcd0d13fe80  ./p0/baked-constrained-pack-build.json
+bf514ce9a0aef75c28ac55ef6243eb1043aa67a733c1e864197b9d493d8c1171  ./p0/cold-restage-cleanup-receipt.json
+9271af0f90160566729a6507f8b8e07ac7d05699e122701110393035b6135f26  ./p0/cold-restage-memory-trace.tsv
+1cbbd4f5e177a16ff3de4a179fc5a8763a37c1d410266b5eb0155da731527a39  ./p0/cold-restage-summary.json
+59403aa1b8d2a7ed2294a9d7b78566ef5320316e1cdaeff67e5d3d06bc031d35  ./p0/frozen-rc-memory-trace.tsv
+aba59dd8e1a8051e4ea93400e0e4931e958c75d239c7dd56444afa7136eee710  ./p0/frozen-rc-summary.json
+005b225341e1303db4560c782d09a6eb8c9422c12267d74712c381fb71e33add  ./p1/aggregate.json
+aa572d1bfa31827f949bef0e422e6593bc4889472cbf6d2ed53b70e3cd84e74a  ./p1/cleanup.json
+0aa4e2f999a721c0a8bd36521cc3487c5b4297958de8b02abfd63a2b992450cb  ./p1/context-30k.manifest.json
+cd2c7621f168683aee68e37c9008dd5faecc808b74fe93ed3b39405f841393fe  ./p1/context-30k.receipts.jsonl
+1afb9b0d96d05c9d7a374dff88c94c76a3e854c8d046fb8ef1783275ad3f5630  ./p1/pool-gate-policy.json
+e32564fb10f3302538af4d8825f3ec95637399f1d54d7dd7aac11dfb13d902de  ./p1/runtime-clean.json
+e559c313cbfeb47a79743d6bc73c84f1244ac20ed36d85d07e49587bfae110fb  ./p1/seed42.manifest.json
+1c4978408014aff023ddb1b71eae0b9f4e8fee59e9cb4434bc38fabd05756487  ./p1/seed42.raw.jsonl
+3c2bdf2e75a0967028a70ca065a0e67c8e70daf4c1ccab1d84491524155f1da0  ./p1/seed42.warmup.jsonl
+b5db436fa14bc8ff3864030212e62c1e5a1ff3217f6eb8fd0ace845220f7ae54  ./p1/seed43.manifest.json
+77cee2bd64cdffa4deb4ac23352cf6053e1bbb02f87bda9b24ab621358d40630  ./p1/seed43.raw.jsonl
+474fb509940ff0a3c143d709538edcec81764d1ea15559f118b8da3fc3a39313  ./p1/seed43.warmup.jsonl
+75c4bbf593ebc84d7d57ea778374f487a6665d2a51d90dd9669597469e0a5dd7  ./p1/seed44.manifest.json
+10a93bf29ff4bd411535f2b48d3d5dfc28ff08f6b6435169b7b505e27c9569ac  ./p1/seed44.raw.jsonl
+f832cca3346094be00650267899a1890c5c40e961fd64d60b203238c1a2c9db8  ./p1/seed44.warmup.jsonl
+5ff7e3d0d22073fdba1f0413aca5c10db573fbf235596746559fee5ec7f22c0f  ./p1/server.json
+5ee72db204c8e30574d426a83782474ac7ec2f67f904da0ad04ff10dab0308a3  ./p1/strict-zero-residue-diagnostic.json
+cad626e05b01464f81f1bb58434e3f6bec5886f69223e8cc8d037dd17d5963ab  ./p2/cleanup.json
+6e16b2af587301748b96d5b24a3e06acff8b73b171b6a483918b2338d63771b3  ./p2/context-120k.manifest.json
+67c2a1f834b9a6ce54d4ea4f2ce98a51c5257bfb8a78cf9a3917bbab52ea1db7  ./p2/context-120k.receipts.jsonl
+db649f7008e2a0bd3a06a51ab242e884b2846eee2f9985ca06ef557ff7a5ac14  ./p2/frozen-rc-sentinel.json
+1afb9b0d96d05c9d7a374dff88c94c76a3e854c8d046fb8ef1783275ad3f5630  ./p2/pool-gate-policy.json
+40362d6150626552760a7843b8ea5454e0d71c0ed1af60f161b6783788f799c9  ./p2/quality-aggregate.json
+fb4732bfa7700598ba8686d8d585c7508848b215bad690ba6e25a4a97c44fbac  ./p2/readiness.json
+1912d3b9d34aafbbbc72fb0a08212cddbbf492a6e4063567f907321b5917d0a0  ./p2/runtime-clean.json
+1d5222c1195ff47766ca7b2a2ab69dd664acbe8c848f0a893170947c66c780f4  ./p2/seed42.manifest.json
+db5b50a8ed30dd68f61431cae477cf65902768d400b4f7d20fc75790c5abc2b6  ./p2/seed42.raw.jsonl
+c0f30b140c53006c5d2041f7e3f75f7dc977bd8980c059aec745c7ea559a909b  ./p2/seed42.warmup.jsonl
+17996af880a997bb5fd1967845b1da14a6af2188de43b571984d83b20b8bde33  ./p2/seed43.manifest.json
+12ebd8458751bbea4d7683794a5c9e7c447784a93d2fae7dce275d59f242f5fc  ./p2/seed43.raw.jsonl
+c89b523c53ee899193e829eaa1b58c06cf992f1ffb587433f293221cbe9ec86d  ./p2/seed43.warmup.jsonl
+58834e56c51c4ef94cc591ed4eb0c1521e20111a4cc623911b2b76cd58c5b443  ./p2/seed44.manifest.json
+d188e788324b3d47a8a81112e7c1e185e1297e703c78df6f30d19b5c0dae7600  ./p2/seed44.raw.jsonl
+e1485a5efec3960924b722834bd04d87112bf32ad1a98373a9981455412d08ba  ./p2/seed44.warmup.jsonl
+f4514891b1a84f1415b3dafaac76bfa8f1261f0804561201d683faf9f84fd8ef  ./p2/server.json
diff --git a/evidence/public/ds4-w2-2026-07-11/exact-image-verification.json b/evidence/public/ds4-w2-2026-07-11/exact-image-verification.json
new file mode 100644
index 0000000..514bd4a
--- /dev/null
+++ b/evidence/public/ds4-w2-2026-07-11/exact-image-verification.json
@@ -0,0 +1,99 @@
+{
+  "schema_version": "vllm-moet-exact-image-verification-v3",
+  "frozen_release_candidate": {
+    "official_vllm_commit": "ee0da84ab9e04ac7610e28580af62c365e898389",
+    "origin_main_cutoff": "94be3aa3d7a8b82c7fc9687990a7edb6035f69f3",
+    "wrapper_head_commit": "0a73c2bcf262ba0aa53560d6290ad2af350e34fc",
+    "source_patch_sha256": "241ba984b1c56f5dc7adbc8d7f519d60b5746024bf7dfeb875e3546a668e79a7",
+    "image_id": "sha256:fc6e1244d60855fe45ccc0236daaaa722abcb8d354200eb92aca104bd954d3f2",
+    "image_source_patch_label": "241ba984b1c56f5dc7adbc8d7f519d60b5746024bf7dfeb875e354200eb92aca104bd954d3f2",
+    "image_wrapper_commit_label": "0a73c2bcf262ba0aa53560d6290ad2af350e34fc"
+  },
+  "immediately_preceding_comprehensive_artifact": {
+    "official_vllm_commit": "ee0da84ab9e04ac7610e28580af62c365e898389",
+    "integrated_origin_main_commit": "7d2429c109b625f23ed7926eb4e80f5d8d80f2e5",
+    "wrapper_head_commit": "5e69228e1e8e6f7345ae96657ab33b50e5805be7",
+    "source_patch_sha256": "41d7b2f96ca3b966cac1b7ed5cff37bc03c27c616ed254aca961cc75a9ffe31d",
+    "image_id": "sha256:66abc2f145244e03ff0f0fcca088be813ad311f2963c62b281e4bb888ac605e9",
+    "validation_scope": "three quality seeds and three exact 120K context depths"
+  },
+  "applied_tree": {
+    "files_changed": 70,
+    "insertions": 14223,
+    "deletions": 226,
+    "apply_check": "pass",
+    "source_diff_whitespace_check": "pass",
+    "python_compile": "pass",
+    "fresh_applied_source_match": "70/70 baked source files matched"
+  },
+  "baked_source_sha256": {
+    "vllm/forward_context.py": "b3b5d84c16e0f5c886186cc06fcd65e1aa577ee6e58c2746ecea4c599f778ec7",
+    "vllm/v1/worker/gpu_model_runner.py": "63a08d992701cfca677de0e35eab2bc00e97a4514ae6d69858d544460d35be3f",
+    "vllm/v1/worker/gpu_ubatch_wrapper.py": "cc9526d859ccecdeb69b42140c92ba13c6fb80a8091734ab518a819c322dd1e7",
+    "vllm/v1/worker/gpu_worker.py": "76331291a8cd98236a17d85ffced0a4ed2d749da7e68f163e5ffb3df922faca5",
+    "vllm/model_executor/layers/quantization/utils/moe_w2_cubit.py": "3fa37293b68084358bc0bf6c9630a81ee722d2c8a936f010837f71f7cde27ca0",
+    "vllm/model_executor/layers/quantization/utils/moe_w2_delta.py": "25235377b2e86c0a5bc84d7fdbcf7bdac0a57f8b4080a66754e795b902bbe736",
+    "vllm/model_executor/layers/quantization/utils/moe_w2_gate.py": "d035e371fd93a7840733ecb3c9b2b89675a5448efac0da834fd5628c1052691b",
+    "vllm/model_executor/layers/quantization/utils/moe_w2_planes.py": "3555b5561063ca000be91fabaf3e65a8da4b56e6c75e93c46611aa82c66e5a56",
+    "tests/model_executor/layers/quantization/test_moe_w2_padded_routes.py": "5452b492bcdeb64413dab659e5baa1b28571f08af1d6cb2abf67a25836a054d1",
+    "tests/model_executor/layers/quantization/test_moe_w2_step_pins.py": "aed09d21defaba3e524a98640005c9ba00b8b8bc3dac539ec5948250743ea1db"
+  },
+  "canonical_synthetic_safety_probes": {
+    "checkpoint_residency_sha256": "e7eb19fc75dc8b49d587b9f07c87c5db182cd764a2f85998aac78d85a03698b8",
+    "checkpoint_next_shard_retry_sha256": "27de4d773ac034dce3d9f6d29b70275ec195219fd69077343a47843dc764c107",
+    "constrained_pack_build_sha256": "28d4b707ff1f49cf3b52091483a47d20ae855f001bd9937b48a69dcd0d13fe80",
+    "image_identity_match": true
+  },
+  "frozen_rc_exact_image_validation": {
+    "source_focused_manager_window_cgroup_route_pin": "21 passed on CUDA",
+    "store_safety": "37 passed",
+    "mxfp4_stream_safety": "10 passed",
+    "gpu_store_backends": "all pinned, mmap, reboot, tiered, eviction, overflow, scan, and preheat checks passed",
+    "forward_matrix_e8": "base 2-bit, full-FP4 delta, and split-FP4 refinement passed",
+    "three_tier_split_e8": {
+      "mixed_max_relative_error": 0.02446,
+      "mixed_cosine_similarity": 0.999867,
+      "transient_miss_count": 2,
+      "transient_max_relative_error": 0.02661,
+      "eviction_mapped_experts": "4/4",
+      "result": "pass"
+    },
+    "checkpoint_residency_128_mib": "pass",
+    "checkpoint_retry_512_mib": "pass",
+    "constrained_pack_48x8x4_mib": "pass below 2 GiB with zero swap"
+  },
+  "split_fp4_base_cache_boundary": {
+    "support": "split-FP4 refinement over the base cache is supported behind the explicit split flag with residency coupling and the fp4s tier tag",
+    "serving_sentinel": "split mode was disabled, so the guarded release-candidate serving run used full-nibble FP4 delta planes",
+    "separate_fixture": "the exact-image E=8 three-tier fixture covered mixed residency, a transient base miss, and eviction"
+  },
+  "serving_validation": {
+    "frozen_release_candidate": {
+      "quality_seeds": [
+        42,
+        43
+      ],
+      "responses": 80,
+      "strict_machine_exact": 79,
+      "semantically_correct": 80,
+      "sink_detections": 0,
+      "retry_free_prewarms": "20/20",
+      "pool_gates": "4/4",
+      "exact_120k_context": "1/1 at depth 0.9 with zero prompt-token tolerance"
+    },
+    "immediately_preceding_comprehensive_artifact": {
+      "quality": "120/120 machine-exact, 120/120 semantically correct, 0/120 sinks across three seeds",
+      "retry_free_prewarms": "30/30",
+      "pool_gates": "6/6",
+      "exact_120k_context": "3/3 at depths 0.1, 0.5, and 0.9 with zero prompt-token tolerance"
+    }
+  },
+  "project_validation": {
+    "evaluation_and_context_harness": "39 passed, 17 subtests passed",
+    "ruff": "changed W2 harness and targeted project files pass; broad upstream helper lint is outside this receipt",
+    "format": "changed W2 harness and targeted project files pass",
+    "benchmark_lint": "0 errors, 0 warnings",
+    "benchmark_render_check": "up to date"
+  },
+  "evidence_boundary": "The frozen release candidate is the current exact image and received focused exact-image GPU, cold-restage, two-seed quality, and deepest-context sentinel validation. The immediately preceding 41d7b2f artifact remains the comprehensive three-seed and three-depth serving series; those broader counts are historical evidence and are not relabeled as frozen-RC reruns."
+}