fix+chore: 30+ fixes bundle F — deferred criticals + MCP+round5 bulk by lusoris · Pull Request #859 · VMAFx/vmafx

lusoris · 2026-06-08T19:34:25Z

Summary

Mega-bundle of ~30 fixes across three rounds (deferred criticals, MCP test suite, round-5 bulk hunt):

Deferred criticals (rounds 4-5 skip-pile, 8 commits):

HIP WARP_SIZE replaced with runtime warpSize across 10 kernel files for gfx10+/wave32 correctness
vmaf_tiny_v2/v3/v4 double-scaling fix via onnx_has_scaler sidecar flag; C-side scaler skipped when ONNX has one baked in
gpu_picture_pool partial-alloc rollback: extracted alloc_pictures() helper with mutex + slot cleanup on failure
vmaf_use_feature_extractor / vmaf_use_features_from_model NULL-deref guard: *fex_ctx=NULL on all error exits + early-return in caller + destroy on state-setter failure (3 files)
Rust Context<'a> + PhantomData<&'a mut Model> prevents UAF when registered model is freed before context
Rust vmaf_read_pictures FFI: explicit vmaf_picture_unref on rc<0 early-error path to free plane buffers
Docker base image digest-pinning for nvidia/cuda:13.3.0 and golang:1.23-bookworm (4 Dockerfiles)
fr_regressor_v2_ensemble_v1.json SHA256 refresh — 5 stale digests updated to match on-disk ONNX files

MCP test fixes (1 commit, 5 fixes): round2_client env scope, async/sync wrapper, round3 patch scope, REPO path monkeypatch, tempdir cleanup assertion

Round-5 bulk (1 commit, 18 fixes): pdjson SIGABRT depth-overflow guard, thread_pool POSIX while→if spurious-wakeup fix, libvmaf + thread_pool NULL guards, motion_avx512 UBSan cast, opt.h const-correctness, SYCL reqd_sub_group_size, Go pkg/errors migration, Rust assertion fixes, build flags, dead-code removal

Race-fix commit (w7nry8u3o) not included — task output was empty (still in flight).

Test plan

meson test -C build --suite=fast: 87/87 PASS
cargo check -p vmafx: clean
No conflict markers (git grep '^<<<<<<<' clean)
No Netflix golden assertions touched
Pre-commit hooks passed; push hooks ran normally

Deep-dive deliverables (ADR-0108)

Research digest: no digest needed: trivial — root-causes documented per-commit from rounds 4-5 hunt findings
Decision matrix: no alternatives: only-one-way fix per finding (NULL-guard, unref, lifetime param)
AGENTS.md invariant note: no rebase-sensitive invariants introduced
Reproducer / smoke-test command: per-fix verifies in individual commits (meson test --suite=fast, cargo check, ONNX sha256sum cross-check)
changelog.d fragment: no changelog fragment needed: mega-bundle (closes ~20+ T- rows)
docs/rebase-notes.md: no rebase impact: additive fixes, no renamed symbols or moved API surfaces

state.md touch

Closes ~20+ T- rows across deferred criticals, MCP test suite, and round-5 bulk categories

🤖 Generated with Claude Code

…kernels gfx10+ (RDNA2+) devices run in wave32 mode; all HIP kernels that hardcoded WARP_SIZE=64 produced wrong results (incorrect warp-reduce strides, lane detection off by factor 2, warp-partial shared arrays under-allocated on wave32 hardware) because the reduction loops iterated 32 lanes when only 16 were live. Fix pattern applied across 10 kernel files: - Replace `#define *_WARP_SIZE 64` with `*_MIN_WARP_SIZE 32` (for shared-mem sizing — allocates enough slots for wave32 worst case). - Replace `WARPS_PER_BLOCK = BX*BY / WARP_SIZE` denominator with MIN_WARP_SIZE so shared-mem arrays are large enough on wave32 (over-allocated on wave64, which is correct and safe — unused slots are never written). - Replace all `for (off = WARP_SIZE/2; ...)` with `(int)warpSize / 2`. - Replace all `lid % WARP_SIZE` / `lid / WARP_SIZE` with `% (unsigned)warpSize` / `/ (unsigned)warpSize`. - Replace compile-time `WARPS_PER_BLOCK` loop bounds in the final accumulation loops with runtime `nwarps = blockDim.x * blockDim.y / warpSize` so wave64 does not read uninitialised shared-mem slots 2-3 that wave32 would write. Files changed (9 kernels + 2 host launchers): float_adm_score.hip, float_ssim/ssim_score.hip, integer_ciede/ciede_score.hip, integer_moment/moment_score.hip, integer_motion/motion_score.hip, integer_motion_v2/motion_v2_score.hip, integer_ms_ssim/ms_ssim_score.hip, integer_psnr/psnr_score.hip, integer_ssim/integer_ssim_score.hip. speed/speed_score.hip: HIP_WARP_SIZE is a launch-time parameter (blockDim.x must equal the device wavefront size). Updated speed_chroma_hip.c and speed_temporal_hip.c to query hipDeviceProp_t.warpSize at init and store it in state->solve_warp, replacing the hardcoded SC_SOLVE_WARP/ST_SOLVE_WARP constant at hipModuleLaunchKernel call sites. integer_vif/vif_statistics.hip was already fixed in PR #850 (ADR-0563 per-thread atomicAdd); float_motion, float_psnr, float_vif, float_moment were already using warpSize at runtime. Local verify: meson setup -Denable_hip=true + ninja 1087/1087 OK. The 9 fast+gpu HIP parity failures are pre-existing on master (confirmed by baseline run on origin/master before this change). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

The vmaf_tiny_v2/v3/v4 ONNX graphs bake the StandardScaler as Constant nodes (ADR-0244), so the runtime should feed raw feature values. However, the sidecar JSON files echo input_mean/input_std for tooling, and vmaf_dnn_sidecar_load sets has_feature_scaler=true whenever it finds those arrays. The C runtime in vmaf_ctx_dnn_run_frame_feature_vector then unconditionally applied (v - mean) / std before calling ORT — effectively double-scaling every feature vector and corrupting vmaf_tiny_v{2,3,4} scores. Fix: add an onnx_has_scaler boolean field to VmafModelSidecar and the sidecar JSON spec. When true, the C runtime skips its own scaler step because the ONNX graph handles normalisation internally. The three affected sidecars (vmaf_tiny_v2.json, v3.json, v4.json) are updated to set "onnx_has_scaler": true. Older models without this flag (fr_regressor v1/v2/v3 etc.) are unaffected — their ONNX graphs do NOT bake the scaler, so the C runtime continues to apply it. Two new unit tests are added to test_model_loader.c: - test_sidecar_onnx_has_scaler_flag: verifies parsing onnx_has_scaler while has_feature_scaler remains true for tooling. - test_sidecar_onnx_has_scaler_absent: verifies backward compatibility (flag defaults to false when key is absent from the sidecar). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… failure vmaf_gpu_picture_pool_init accumulated errors with |= but never freed pictures that were successfully allocated before the first failure, leaking one GPU buffer per successfully-completed slot. Extract alloc_pictures() static helper that breaks on first failure and rolls back slots 0..alloc_cnt-1 via free_picture_callback before returning the error. Also call pthread_mutex_destroy on the alloc-failure path so the mutex does not leak its internal state. The helper also brings vmaf_gpu_picture_pool_init back under the 60-line readability-function-size threshold (was 64 lines). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ate failure vmaf_feature_extractor_context_create freed and returned *fex_ctx as a dangling pointer on error (malloc failure or vmaf_fex_ctx_parse_options failure), without zeroing the out-parameter. Two call sites in libvmaf.c then used err |= set_fex_cuda_state(fex_ctx, ...) unconditionally, dereferencing the freed pointer before the err-check. Fix in three parts: 1. feature_extractor.c / .cpp: set *fex_ctx = NULL on all error exits so callers always receive NULL on failure, never a dangling pointer. 2. libvmaf.c (vmaf_use_feature_extractor): early-return on create failure before any set_fex_*() call; also destroy fex_ctx on state-setter failure. 3. libvmaf.c (vmaf_use_features_from_model): same pattern. The pool slot allocator (ctx_pool_claim_slot) and predict.c already handled the error correctly; no change needed there. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

vmaf_use_features_from_model stores the raw VmafModel* inside the context's feature-collector linked list (vmaf_feature_collector_mount_model in core/src/feature/feature_collector.c, line 272: m->model = model). Libvmaf dereferences that pointer during score computation, so if the Rust Model wrapper was dropped while the Context was still alive, every subsequent vmaf_score_pooled / vmaf_score_at_index call would be a use-after-free. Add lifetime parameter 'a to Context<'a> with PhantomData<&'a mut Model> (invariant over 'a to prevent coercion to a longer lifetime), and bind use_features_from_model to model: &'a mut Model. A freshly-constructed Context<'static> carries no borrow; the borrow checker narrows the context's apparent lifetime to the shortest-lived registered model as soon as use_features_from_model is called, preventing the context from outliving any registered model. ContextBuilder::build() explicitly returns Result<Context<'static>>. Local-verify: cargo check + cargo clippy -D warnings, both clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

When `vmaf_read_pictures` returns an error, it has not taken ownership of the plane buffers. The Rust-side `Picture` wrappers already had their `owned` flag cleared by `into_raw_owned()`, so their `Drop` became a no-op. The resulting raw `VmafPicture` structs then fell out of scope without being freed — a silent memory leak on every rejected frame. Fix: check `rc < 0` immediately after the FFI call and explicitly call `vmaf_picture_unref` on both raw pictures before returning the error. Added a unit test (`into_raw_owned_clears_owned_flag_and_caller_frees`) in `picture.rs` that exercises the same manual-unref path. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

@sha256

…ase images Four Dockerfiles referenced golang:1.23-bookworm (controller, node, operator) and nvidia/cuda:13.3.0-devel-ubuntu24.04 (root Dockerfile) without @sha256 digest anchors, leaving supply-chain provenance unverifiable and builds non-reproducible across rebuilds. Pin each with the current manifest digest so BuildKit verifies the layer before pulling. Also removes the stale TODO comment in Dockerfile that deferred the nvidia/cuda pin until the digest "stabilises". - Dockerfile: nvidia/cuda:13.3.0-devel-ubuntu24.04@sha256:ef220... - docker/Dockerfile.controller: golang:1.23-bookworm@sha256:16705... - docker/Dockerfile.node: golang:1.23-bookworm@sha256:16705... - docker/Dockerfile.operator: golang:1.23-bookworm@sha256:16705... Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… manifest The five per-member sha256 fields in the ensemble manifest were out of sync with the actual ONNX files on disk. The per-seed sidecar JSONs already carry the correct hashes (verified); only the parent manifest was stale. Updated fields (old → new): seed0: fa89df1b... → 08ab1aed... seed1: 48b0b9f3... → 42d86249... seed2: 927a23fe... → 0f1d8a81... seed3: 1be056b0... → ed5e9260... seed4: 584ad69b... → 036c6224... Verification: sha256sum of each .onnx file matches the updated manifest entry. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

1. round2_client fixture: extend VMAFX_MCP_HTTP_NO_AUTH patch to cover the full test duration (yield inside with-block) so the security middleware sees the bypass at request-dispatch time, not just at app-creation time. 2. test_vmaf_version_handles_version_timeout: drive async _vmaf_version() via asyncio.run() instead of calling it synchronously (ADR-1023 made the function async). 3. test_coverage_round3.py HTTP tests: extend patch.dict scope to wrap client.post() + assertions in all three tests; the with-block was exiting before the request was made. 4. test_describe_model_onnx_no_metadata: replace hardcoded REPO constant with monkeypatch.setattr(srv, '_repo_root', lambda: tmp_path) to decouple from the physical install layout. 5. test_describe_worst_frames_uses_unique_tempdir_per_call: update assertion to match TemporaryDirectory semantics (files cleaned up on context exit — assert NOT exists, not exists). All 8 previously-failing tests now pass (0.44 s). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…Rust/warnings bulk 18 files, 19 distinct fixes across 11 of the 12 round-5 hunt cells (race-conditions cell handled separately): memory-leaks-strict (cell 2): - core/src/libvmaf.c: set *vmaf=NULL after free(v) in free_v: path to prevent dangling pointer after vmaf_init sub-init failure - core/src/thread_pool.c: set *pool=NULL after free(p) on workers-malloc failure; change if→while for pthread_cond_wait to guard spurious wakeups - core/src/feature/feature_extractor.c: set *pool=NULL after free(p) at free_p: label; add comment documenting intentional mutex-init ordering tiny-ai-model-runtime-audit (cell 9): - core/src/pdjson.c: fix UBSan SIGABRT — change stack_top=-1 to (size_t)-1 at both init sites, matching the (size_t)-1 sentinel used at the two comparison sites (lines 720/746) cross-extractor-numeric-comparison (cell 5): - core/src/feature/x86/motion_avx512.c: fix signed integer overflow in three 16-bit scalar-tail paths (y_conv_8, y_conv_16, x_conv_16) — cast uint16_t filter[k] and src pixels to uint32_t before multiply so products stay in unsigned arithmetic (UBSan SIGABRT fix) hidden-config-options (cell 6): - core/meson_options.txt: fix inverted enable_nvcc description (was "Use clang…"; corrected to "Use nvcc…; when false, clang is used") - docs/development/build-flags.md: remove false claim that enable_vulkan remains as a failing stub; the option() was fully deleted per ADR-0726 - core/src/meson.build: remove dead -DOC_NEW_STYLE_INCLUDES legacy define (no C consumer in tree since upstream libogg era) - core/src/feature/adm.c: collapse ADM_OPT_SINGLE_PRECISION #ifdef branch (symbol never defined anywhere; 1e-2 threshold was permanently dead code) unused-code-dead-code (cell 7): - core/src/picture.h: remove dead VMAF_PICTURE_BUFFER_TYPE_VULKAN_DEVICE enum value (Vulkan removed per ADR-0726; no reference sites remain) meson-build-warnings-grind (cell 12): - core/src/opt.h: char*→const char* for default_val.s to prevent write-to-string-literal UB (icpx -Wwritable-strings) - core/src/opt.c + opt.cpp: propagate const through set_option_string signature; opt.cpp uses const_cast per ADR-0721 ABI-stability comment - core/src/feature/sycl/sycl_compat.h: update VMAF_SYCL_REQD_SG_SIZE macro from deprecated [[intel::reqd_sub_group_size(N)]] to SYCL 2020 standard [[sycl::reqd_sub_group_size(N)]] (oneAPI 2026.0 deprecation warning) go-pkg-error-handling-audit (cell 10): - pkg/libvmaf/libvmaf.go: check tmpOut.Close() error; use logged deferred remove instead of bare defer os.Remove - cmd/vmafx-controller/queue/queue.go: surface json.Marshal error in ReportResult instead of silently falling back to "{}"; fix ListAll defer rows.Close() to log close errors (matching reload() pattern) - cmd/vmafx-mcp/impl.go: intArg/floatArg return def (not 0) on json.Number conversion failure rust-bindings-safety (cell 11): - bindings/rust/vmafx-sys/tests/integration_test.rs: add debug_assert for non-negative strides before as-usize cast; add allow(cast_sign_loss) with justification comment Build verified: ninja -C /tmp/vmaf-round5-build (CPU-only) 894/894 OK, meson test --suite=fast 87/87 OK (including test_motion_avx512_parity which previously aborted under UBSan). go build clean. cargo check clean. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…RF, AI scripts, MCP conformance) (#865) * fix(hip,sycl): stale wave32 comment + Kahan IIR blur for ssimulacra2 (iter6-cross-backend-parity) Three iter6 cross-backend-parity findings: 1. [critical — partial] float_adm_score.hip: wave32 code was already fixed in #859 (0a9dba8); correct the lingering stale comment that still read "FADM_WARPS_PER_BLOCK = 4 (64-lane warps)" — FADM_WARPS_PER_BLOCK is now 256/32 = 8 slots (sized for the wave32 worst case). No functional change. 2. [high — already fixed] float_ssim/ssim_score.hip + integer_psnr/psnr_score.hip: wave32 runtime-warpSize fixes landed in #859 alongside float_adm; nothing further to do here. 3. [high] ssimulacra2_sycl.cpp: add Kahan (compensated-summation) state tracking to the 3-pole recursive IIR blur kernel (launch_blur<PASS>). The IIR state (prev1_k) accumulated O(eps) rounding error per step; over 4K-tall frames this exceeded the 5e-5 cross-backend parity contract vs the CPU reference. Each pole now carries a float comp_k compensation term; the standard Kahan pattern (y = candidate - comp; new_state = old_state + y; comp = (new_state - old_state) - y) bounds per-iteration error to O(eps^2) without fp64 (ADR-0220 compliant). No Netflix golden-data assertions modified. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ubsan): replace 0xFFFFFFFFFFFFFFFF hex literals with -1LL in adm_avx512.c _mm512_set_epi64 takes long long (signed 64-bit) arguments. The literal 0xFFFFFFFFFFFFFFFF exceeds LLONG_MAX and is undefined behaviour under strict UBSan. Replace with -1LL which has identical bit pattern and correct type at both call sites (ADM_CM_THRESH_S_I_END macro lines 406 and 698). Finding: iter6-ubsan-strict high [adm_avx512: 0xFFFF... overflows long long]. Note: motion_avx512.c and adm_avx2.c analogous fixes were already present on master (PR #858 / PR #859 bundles). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(thread-safety): eliminate 2 TSAN races in threaded batch path (iter6-tsan-race-deep) Finding #1 (high): Remove the racy write `fex->framesync = vmaf->framesync` in `threaded_enqueue_one`. `fex` is the *shared* registered VmafFeatureExtractor — worker threads from previous frames may concurrently read its fields. The write is redundant: framesync is already propagated to every pool-slot copy by `set_fex_framesync()` at registration time and by `ctx_pool_ensure_slot_ctx()`. Finding #2 (high): Move the `vmaf->prev_ref` advance to BEFORE the enqueue call in `threaded_read_pictures_batch`. In the old order the main thread unreffed and replaced `vmaf->prev_ref` after enqueue while the just-submitted worker still held a live reference to the same underlying VmafRef*, creating a concurrent unref/write on the same object without synchronisation. Workers use `data.prev_ref` (an independently refcounted snapshot) exclusively and never re-read `vmaf->prev_ref` after enqueue, so moving the advance before enqueue is both safe and race-free. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(vmaf-tune): fix --no-bisect TypeError and recommend CRF selection strategy Two bugs in the compare/recommend paths: 1. _encode_and_score() in bisect.py had encode_runner and score_runner as required keyword-only args (no defaults). The CRF-sweep caller in cli.py did not pass them, causing an unconditional TypeError on any compare --no-bisect --crf-sweep invocation. Fixed by giving both parameters a default of None, consistent with the existing decode_runner parameter and the run_encode/run_score runner=None semantics. 2. _smallest_passing_crf() in cli.py used `crf > cur[0]` to select the largest (most efficient) passing CRF, contradicting both the function name and the CLI help string ("find the smallest CRF whose VMAF >= --target-vmaf"). The correct strategy is the smallest (highest-quality) passing CRF. Fixed comparison to `crf < cur[0]` and updated the docstring to match. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ai-scripts): 3 iter6 runtime bugs — vulkan-device AttributeError, saliency parity always-fail, ensemble ONNX codec-dim mismatch - collect_gpu_calibration_data.py: remove dead args.vulkan_device reference from devices dict (Vulkan removed per ADR-0726; no --vulkan-device argparse registration existed, causing AttributeError at runtime) - validate_saliency_student.py: replace broken PT-reconstruction parity check with ORT-only sanity check when no PT state provided. do_constant_folding=True folds BN stats into conv weights at export so ONNX initializer names diverge from PT state_dict keys; the old code silently left 60 of 65 weights at random defaults and always failed. When pt_state is provided (trainer path) full PT<->ORT diff is still performed. - model/tiny/fr_regressor_v2_ensemble_v1_seed{0..4}.onnx: regenerate with codec_onehot=[batch,6] matching current CODEC_VOCAB (was [batch,14] from a 12-entry encoder_vocab + 2 norm dims; CODEC_VOCAB was later trimmed to 6). Regenerated via train_fr_regressor_v2_ensemble.py --smoke in vmaf-dev-mcp container. eval_probabilistic_proxy.py --smoke now passes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(mcp): JSON-RPC parse-error conformance + deep-nesting guard Two iter6 fuzz conformance findings fixed: 1. [high] Parse-error notification instead of error response (id=null): Add `_ParseErrorFilteredStdin` — a lazy async stdin wrapper that pre-validates each incoming line with `JSONRPCMessage.model_validate_json`. On failure it calls `_emit_parse_error` which writes `{"jsonrpc":"2.0","id":<recovered_or_null>,"error":{"code":-32700, "message":"Parse error"}}` to stdout synchronously, then drops the line so the mcp library never sees a bare Exception on the stream (the notification path is bypassed entirely). `_run()` now passes `stdin=_ParseErrorFilteredStdin()` to `stdio_server`. 2. [high] 500-level deep nesting triggers recursion-limit exception: Add `_check_depth(obj, max_depth=50)` helper and call it at the top of `_call_tool_dispatch` before the tool dispatch. Payloads exceeding 50 nesting levels raise `ValueError`, which the mcp library converts to an isError=True tool result before the pydantic parser recurses. Existing tests updated: the two `stdio_server`-patching tests in `test_coverage_round2.py` now accept `**kwargs` so they tolerate the new `stdin=` keyword argument forwarded by `_run()`. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

bindings/rust/vmafx has existed since ADR-0929 (PR #859); the page still said a higher-level crate was planned for a future PR. Found by the 2026-09-04 gap triage (#1270), verdict STALE, verified against master. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…1282) * docs(dev): replace the impossible venv recipe with the verified one Replace the impossible repo-root pip install -e ".[dev]" in docs/development/languages.md with the verified per-package editable install recipe. The repository root pyproject.toml contains only tool configuration for vmaf-fork-tooling and lacks build-system and root project dependencies, causing flat-layout discovery failure. Per-package editable installs are the model across independent distributions, with dev/Containerfile:1037-1047 as the authoritative reference. Pin meson==1.12.0 to prevent breaking ninja regeneration. Add recovery instructions for virtualenvs broken by legacy tracked .venv symlink loops (PR #1280). docs/state.md: no bug row needed (docs-only). docs/rebase-notes.md: no rebase impact: docs/development/ is fork-added. * docs(dev): rust.md no longer calls the shipped vmafx crate 'planned' bindings/rust/vmafx has existed since ADR-0929 (PR #859); the page still said a higher-level crate was planned for a future PR. Found by the 2026-09-04 gap triage (#1270), verdict STALE, verified against master. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * docs(changelog): rename the venv-recipe fragment so the tracked-venv gate stops matching it scripts/ci/check-no-tracked-venv.sh matches any basename starting with 'venv' (the dot is optional in its pattern); the fragment name venv-recipe-docs.md tripped the required Pre-Commit check on #1282. The gate itself is tightened in a separate PR. --------- Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

lusoris and others added 10 commits June 12, 2026 19:21

lusoris marked this pull request as ready for review June 12, 2026 17:38

lusoris force-pushed the chore/bundle-f-30-fixes-deferred-mcp-bulk-race branch from 2040251 to b3d34cc Compare June 12, 2026 17:38

lusoris merged commit 0a9dba8 into master Jun 12, 2026

lusoris deleted the chore/bundle-f-30-fixes-deferred-mcp-bulk-race branch June 12, 2026 17:38

lusoris mentioned this pull request Jun 12, 2026

fix: 5 Fable deep-hunt findings — integer_ssim AVX2 16-bit overflow, bpc validation, set_fex UAF, vmafx-server DoS cap, CUDA prev_ref #866

Merged

9 tasks done

lusoris added this to the 1.0.0 — First release milestone Sep 4, 2026

lusoris mentioned this pull request Sep 4, 2026

RC9 — Complete tiny-AI training and validation #1242

Closed

6 of 7 tasks

lusoris mentioned this pull request Sep 4, 2026

docs(dev): replace the impossible venv recipe with the verified one #1282

Merged

10 of 13 tasks

Navigation Menu

Repository navigation

fix+chore: 30+ fixes bundle F — deferred criticals + MCP+round5 bulk - #859

fix+chore: 30+ fixes bundle F — deferred criticals + MCP+round5 bulk#859
lusoris merged 10 commits into
masterfrom
chore/bundle-f-30-fixes-deferred-mcp-bulk-race

lusoris commented Jun 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

lusoris commented Jun 8, 2026

Summary

Test plan

Deep-dive deliverables (ADR-0108)

state.md touch

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant