feat: distill diffusion loader and storage boundaries

docs/reference-matrix.md

@@ -45,7 +45,7 @@ Reading rules:
 | NVlabs/kda @ `acdd79343c5e9d0c7cb96289f26fd3f1051ce785` → mit-han-lab/KernelWiki gitlink @ `76d27b56f804e7e7295d4c570e1e5d7eef4b0a75` | Direct bounded distillation, 2026-10-23; cutoff 2026-04-27. All 48 synthesis headers and 2,217 source headers screened; 171 SM90 source headers retained, selected bodies and independently matched DeepGEMM/CUTLASS code plus versioned NVIDIA sections inspected. [Readset/coverage/license qualifications](../src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/references/sources.md); remaining PR bodies/algorithms unmined; KW pin has no root LICENSE, KDA MIT/sanitization assertions qualified | mined (conditional) — Hopper TMA/mbarrier/WGMMA ownership, resource/layout/pipeline/software scheduling and selected FP8 scaling → [reason-hopper-kernels](../src/inferrsi/data/skill_library/hardware/cuda/reason-hopper-kernels/SKILL.md); SM100 recipes and unsafe snippet generalizations excluded; native/precision/performance not_run | out-of-scope — device registers/SMEM/HBM are not host RAM | out-of-scope — no storage implementation extracted | out-of-scope — CTA/cluster synchronization is not network transport | out-of-scope — tile scheduling is not OS scheduling |
 | antirez/h3.c @ `8974cc055ea9c02fcd14cc27dfda3e1027c05153` (MIT; retained ccv BSD-3-Clause notice) | Direct bounded source distillation, 2026-10-23; 34 selected file bytes matched to pinned Git blobs; [readset/architecture](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/references/source-notes.md). Loader, generation/cache, TE/vision/input-output VAEs, AdaLN, selected Metal/int8/SSD paths and native test assertions; full kernels/driver/media internals unread. Prior model-header observations reused, not rerun | mined — joint audio/video schedule and row-map dependencies, BF16 rounding/fusion versus numerical approximation, actual TensorOps dispatch → [audit-h3-runtime-residency](../src/inferrsi/data/skill_library/repositories/h3-c/audit-h3-runtime-residency/SKILL.md); native/quality/performance not_run | mined — shared/file-backed weights, Qwen prefetch ring, phase/cache/preview overlap and complete host output; tensor counters are not physical footprint or 64 GiB qualification | mined — exact raw BF16/F32 schema, direct/mapped reads and two-slot BF16 matrix streaming with completion/join boundaries; uncached-read request is not measured cache bypass or useful overlap | out-of-scope — local GPU/I/O ordering is not network transport | out-of-scope — no OS scheduling/control implementation extracted |
 | Comfy-Org/ComfyUI @ `b1a247c9cf9dbfda4b2c0a0c1d50950667eeaa0c` (GNU GPL v3) | Source-reading archive, inspected 2026-10-21; bounded execution/cache/graph/model-management/static-dynamic-patcher/sampler paths and selected tests; [readset](../src/inferrsi/data/skill_library/repositories/comfy-ui/references/repo-provenance.md); not full-model/custom-node/provider-native coverage | mined — managed weight/patch identity, partial unload versus detach and dynamic-backend boundary → [comfy-ui](../src/inferrsi/data/skill_library/repositories/comfy-ui/SKILL.md); device behavior unverified | mined — output versus object identity, ancestor fingerprints, lazy consumer references and retention versus validity; host/pin ownership | out-of-scope — no storage implementation extracted | out-of-scope — no transport implementation extracted | out-of-scope — workflow scheduling is not OS scheduling |
-| leejet/stable-diffusion.cpp @ `b13a48219d5cdf408edcb93da6cd69b612ed03d3` (MIT) | Supplied fixed-revision archive, inspected 2026-10-22; 18 bounded readset paths plus one separately identity-checked README matched to pinned commit/tree (19 blobs total); header coverage is selected owner/state declarations; full archive/submodules not byte-authenticated; bounded image API, native runner, capacity, segment/prefetch and cleanup paths; [readset](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/source-notes.md); GGML submodules/model kernels unread | mined — native weight/runtime capacity, protected segment/prefetch lifetime → [audit-sd-cpp-residency](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md); native/backend behavior unverified | mined — parameter storage versus compute staging and retained runtime ownership; host peaks unmeasured | out-of-scope — demotion eligibility observed, complete loader/I/O not extracted | out-of-scope — device copies are not network transport | out-of-scope — runner scheduling is not OS scheduling |
+| leejet/stable-diffusion.cpp @ `b13a48219d5cdf408edcb93da6cd69b612ed03d3` (MIT) | Supplied archive, inspected 2026-10-22: 18 bounded paths plus separately identity-checked README (19 blobs); historical receipt preserved. 2026-10-23 extension: 12 selected C++ blobs plus retained MIT text identity-checked, with line-scoped catalog/non-ZIP safetensors read/staging/mapping and immediate consumer cleanup; [readsets](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/source-notes.md). File identity is not full-file review; archive/submodules, model kernels and other-format implementations remain outside the bounded audit | mined — native weight/runtime capacity, protected segment/prefetch lifetime → [audit-sd-cpp-residency](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md); native/backend behavior unverified | mined — parameter storage versus compute staging and retained runtime ownership; loader host read/conversion buffers versus mapped owners; host peaks unmeasured | mined — bounded catalog stamps versus content identity, range admission versus actual reads, mapping/shared-owner and newly created block cleanup → [loader/storage boundaries](../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/loader-storage-boundaries.md); full other-format parsing, native failure/SSD/performance qualification remain open | out-of-scope — device copies are not network transport | out-of-scope — runner scheduling is not OS scheduling |
 | NVIDIA/TensorRT-LLM @ f33e0e5fd | skill-doc-only (`.claude/skills/`) | mined — evidence taxonomy, kernel writing, casebook, roofline refs → `dense-performance-evidence`, `kernel-writing-guide`, `inference-optimization-casebook`, `performance-ceiling-analysis` | mined — host overhead verdict + line-profiling loop → `host-overhead-profiling`, `host-overhead-optimization` | mined — weight-movement/overlap + KV host-cache patterns → `weight-loading-optimization`; disk-level: negative (see below) | out-of-scope — host/device weight-movement overlap remains in RAM/Storage, not network coverage | mined — CPU affinity rule + env preflight → `os-process-placement` |
 | Ascend/MindIE-SD @ 8481d39 | skill-doc-only (`.agents/skills/`) | mined — 5-layer analysis, bottleneck casebook → `ascend-performance-analysis`, `ascend-optimization-casebook` | mined — host-opt async/pattern sections → `weight-loading-optimization` | mined — load-path judgment, fixed-overhead accounting, offload tiers, weights-prep → `weight-loading-optimization` | mined — USP strategy, masking formula, topology rules → `ascend-distributed-parallel` | mined — device health, card-group contracts, measurement discipline → `ascend-device-health` |
 | ascend-ai-coding/awesome-ascend-skills @ 75ec89c | full (`skills/`, `external/cannbot/`); incremental slice pinned to `75ec89ce9d401be8ac95980e1c3f42f8d4315057` | mined — graph mode, operator profiling/tuning, tiling design → `ascend-graph-mode`, `ascend-operator-profiling`, `ascend-operator-tuning`, `ascendc-tiling-design` | assessed/deferred — KV overview and async-stream recipe lack a complete transfer/completion implementation and retain cannbot license UNKNOWN; [boundary/readset](reviews/issue91-awesome-ascend.md#host-kv-boundary); no new RAM capability | mined — device-side prefetch + diffusers CPU-offload playbook → `weight-loading-optimization`; disk-level: negative | mined (wrapper-source only) — native/PyTorch test selection, tensor-byte meaning, timing/oracle coverage and rank-map counterexamples → `ascend-distributed-parallel`; [full-pin readset](../src/inferrsi/data/skill_library/hardware/ascend/ascend-distributed-parallel/references/awesome-ascend-communication-tests.md); no native-collective or transport audit | mined — env checks, device-busy preflight, thread/env tables → `os-process-placement` |

docs/reviews/issue91-sd-cpp-loader-storage.md

@@ -0,0 +1,83 @@
+# Stable-diffusion.cpp catalog-to-storage construction
+
+This bounded Issue #91 slice extends the existing
+[`audit-sd-cpp-residency`](../../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/SKILL.md)
+candidate. It preserves its stable ID, eight residency steps, MIT text, original
+source pin and historical archive receipt. No engine, dependency, evaluator or
+hardware execution is introduced. Material author: Codex / `gpt-6.1-sol`,
+reasoning effort `xhigh`.
+
+## Scope → ground → construct → verify
+
+| AREX phase | Decision and native InferRSI asset |
+|---|---|
+| Scope | Fill the catalog/source → actual byte delivery → immediate storage-consumer gap. Select ordinary non-ZIP safetensors and its shard index; leave generic startup optimization with `weight-loading-optimization`. Do not replace request-time residency guidance. |
+| Ground | Fixed leejet/stable-diffusion.cpp `b13a48219d5cdf408edcb93da6cd69b612ed03d3`, tree `df847d004c6529eefb02bb629d7ee26e740121a9`. Twelve C++ blobs and retained MIT text have exact byte identity checks; semantic ranges are in the [new ledger](../../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/loader-storage-boundaries.md#fixed-readset-and-attribution). The prior nineteen-blob receipt remains historical. |
+| Construct | Reuse one owner and the native metadata/readers. Add [two loader/storage decisions](../../src/inferrsi/data/skill_library/repositories/stable-diffusion-cpp/audit-sd-cpp-residency/references/loader-storage-boundaries.md), a root route, additive source/validation notes and scoped entries in both matrices. No new ID, backend, counter or parallel registry is needed. |
+| Verify | Check actual owner/resource discovery, metadata, local links, immutable source ranges and retained license. Preserve the old eight-step body and source identity. Bind generated records to a real source commit only after source freeze. Exact execution and independent-review receipts belong to the PR; package/schema checks do not qualify native loading or agent efficacy. |
+
+AREX portable distill/extend guidance is used from fixed
+`ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6`; this is an InferRSI candidate extension,
+not a managed-skill installation. The root license is MIT, Copyright (c) 2023
+leejet, retained as the complete existing owner `LICENSE` (Git blob
+`56e1e5a63852c3bcd5ac4fcdfecb7666f5274169`). No runtime implementation is copied;
+this attribution does not license models, unexpanded dependencies, external ZIP
+code or unrelated embedded helpers.
+
+```mermaid
+flowchart LR
+  S["Fixed engine and loader source"] --> C["Catalog stamps and candidate admission"]
+  C --> B["Actual reads / mapping / host staging"]
+  B --> M["ModelManager completion and owned cleanup"]
+  C --> R["Owner-local loader/storage reference"]
+  M --> R
+  R --> O["Existing residency owner and metadata"]
+  O --> A["Native registry, readers and two matrices"]
+```
+
+## Observations, deductions and execution boundary
+
+Source observation: dependency checks compare path, size and mtime; internal
+source/catalog revisions are generations. Metadata parsing and non-ZIP byte
+bounds precede actual reads. Regular stream reads and mapping copies can return
+failure; host staging/conversion and backend-copy calls are distinct. The
+manager's destination callback records a name before bytes are delivered, while
+loaded-state publication requires all selected source groups to succeed.
+Failure cleanup releases the parameter blocks created for that load. Shared
+mapping owners can remain in loader caches or other blocks.
+
+Deduced audit rule: neither a matching stamp nor a successful destination
+callback is a byte-identity/completion oracle. Require an independent content
+manifest/output oracle when that conclusion matters. Do not infer global
+rollback, pinned allocation, physical SSD traffic, cache eviction or device
+completion from these interfaces. Loader stage counters divided by the last
+file's thread count are not a cold-storage bandwidth denominator.
+
+Native parser/load/failure injection, allocation/unmap behavior, backend-copy
+completion, full-model image quality and performance are **not_run**. Matching
+GGML/private headers/build and a legal minimal fixture are prerequisites, not
+replaceable by mocks or extracted formulas. The source-only readset does not
+qualify all GGUF/Torch/ZIP formats, special quantization/conversion correctness,
+video/audio or every backend. Agent efficacy is **not evaluated / uncovered**.
+
+## Proposed regression cases
+
+All rows are proposed; no case below is a runtime result. Expected decisions
+are stated independently of a source keyword or successful reader invocation.
+
+| Case | Input/condition | Independent expected decision / oracle |
+|---|---|---|
+| New identity positive | A required shard has changed under the selected index | Refresh/invalidate before reuse; inspect real source and consumer state plus caller result. |
+| New identity negative | Same path/size/mtime, different independently recorded digest | Do not certify equal bytes from catalog stamps or generation IDs. Preserve the differing digest. |
+| New admission negative | A selected tensor interval exceeds the real file | The real parser/add-file rejects that candidate; its failed admission does not replace the previous accepted catalog. |
+| New read negative | A file becomes short after catalog admission | The caller observes failure; callback invocation does not count as load completion. Inspect loaded flags/new-block cleanup without presuming universal rollback. |
+| New staging positive | Supported ordinary dtype with the actual read/staging/mapping path selected | Compare real tensor bytes to an independent legal small-fixture oracle. Keep input/target byte counts separate and owners alive through use. |
+| New lifetime negative | One block releases its map owners while another owner/cache retains the mapping | Do not predict final unmap or memory savings from one release. Check actual retained owners and final destruction. |
+| Old dispatch regression | CPU or multi-device scheduling with a small `max_vram` | Preserve the existing distinction from the non-CPU single-device segmented path; a small budget alone is insufficient dispatch evidence. |
+| Old lifetime regression | An active/queued weight becomes unpinned | Preserve the existing distinction between eligibility for reclamation and immediate physical free; async capability alone is not useful overlap. |
+
+If native prerequisites are later supplied, run the real selected parser/manager
+for the new cases and the actual backend/model for residency/image claims.
+Package/readers validate retained content and access only. Keep preparation,
+first image, warm requests, byte equality/quality, host/device peaks and timing
+denominators separate; no result in one category promotes the candidate.