Fix restricted-pointer cooperative handoffs and expand regressions

.github/scripts/resource_canary_baseline.json

@@ -189,7 +189,7 @@
         "stack": 0
       },
       "canary_block_trsv_f32": {
-        "regs": 17,
+        "regs": 19,
         "smem": 288,
         "spill_loads": 0,
         "spill_stores": 0,

.github/workflows/gh-pages.yml

@@ -101,4 +101,4 @@ jobs:
     steps:
       - name: Deploy to GitHub Pages
         id: deployment
-        uses: actions/deploy-pages@cd2ce8fcbc39b97be8ca5fce6e763baed58fa128 # v5.0.0
+        uses: actions/deploy-pages@368f82528645a54fb793d4d04e342629a3f51346 # v5.0.1

CHANGELOG.md

@@ -66,6 +66,15 @@ All notable changes to GLASS will be documented here. The format follows
 
 ### Fixed
 
+- Restricted optimized callers now reload cross-thread handoffs in triangular
+  solves, softmax/logsumexp, LU pivot selection/permutation, pivoted inverse,
+  pivoted LDLT and symmetric eigendecomposition. This fixes wrong results,
+  invalid accesses and divergent control flow seen despite existing barriers
+  (including issue #44). Thread-only solves retain ordinary loads.
+- Optimized regression coverage now retains the minimal failing caller context
+  alongside independent-oracle checks for shared/global memory, runtime sizes,
+  launch geometry, pivoting and trailing synchronization.
+
 - `glass-dispatch.cuh` is included in the CMake install file list (previously
   a CMake-installed GLASS could not compile the bare measured-default face).
 - CMake propagates the cuBLASDx CUTLASS include path when

CLAUDE.md

@@ -174,7 +174,7 @@ cd docs && PATH="$(cd .. && pwd)/.venv/bin:$PATH" make all SPHINXOPTS="-W --keep
 
 `release.sh` requires a clean `main`, a changelog entry, a complete
 public-overload manifest (every documented overload compile-covered — an
-overload metric, not line coverage), all 21 declared behavioral obligations
+overload metric, not line coverage), all 22 declared behavioral obligations
 passing, a fresh full GPU receipt, and local receipt verification before it
 tags and pushes.
 

docs/agent_debugging_guide.md

@@ -263,8 +263,15 @@ Nasty properties, so check for it deliberately:
   in-place-without-`__restrict__` correct ⇒ a `__restrict__` aliasing miscompile, not your algebra.
   Don't "fix" it by dropping `__restrict__` (you lose the optimization) — kill the shared re-read.
 - **`glass::warp::reduce` is already safe** (shfl-based); `warp::gemm` is flat (no broadcast). The
-  block-scoped `cgrps`/`__syncthreads` twins are safe *today* (full-block fence) but carry a note to
-  prefer `g.shfl(pivot, 0)` if ever called with a warp-tiled group.
+  block-scoped twins also need explicit handoff reads: issue #44 reproduced
+  stale values through restricted callers despite full-block barriers. Use
+  `cooperative_detail::load(bar, ptr)` for a value whose ownership changes
+  between threads; keep the synchronization that orders its producer and
+  consumer. The helper uses a volatile load for cooperating scopes and a
+  normal load for `ThreadBarrier`. It is not an atomic or a substitute for a
+  barrier. Regression coverage lives in `test/test_restrict.py` and compiles
+  at `-O3`, including partial blocks, true runtime sizes, global/shared buffers,
+  multidimensional launches, and immediate consumers after return.
 
 ### 1h. Loop-unroll copies contract FMAs differently → thread-count NON-invariance inside ONE kernel
 

docs/audits/2026-10-28-restrict/CHECKPOINT.txt

@@ -0,0 +1,64 @@
+Implementation complete and full fresh validation passed: 5,126 tests across 8 shards.
+Source commit e537364; final receipt verified with current toolchain and 0.5.0.
+Validated delivery branch: fix-restricted-handoffs.
+See README.txt final section. The pause record below is historical.
+
+RESUMED by explicit user request, 2026-10-28. See README.txt latest implementation entry.
+The following pause checkpoint is historical:
+
+Paused at user's request before implementation edits, 2026-10-28.
+
+Authorization: user wants broad GLASS fixes tonight, full validation, then push
+all completed work. pytest-gpu-proof next release version accepted as 0.5.0.
+Saved improved plan: README.txt in this directory, including execution target.
+No GLASS library/test source edits, commits, pushes, or publications yet.
+Only docs/audits/ is untracked in GLASS. pytest-gpu-proof still has the earlier
+19-file uncommitted Python-3.10/nvcc metadata change, version remains 0.4.1.
+
+IMPORTANT harness defect noticed during implementation preparation:
+extended.cu and derived harnesses contain
+  if constexpr(OP==8||OP==23) glass::block::inv<T,N>(a,s);
+then OP==23 also calls inv_pivoted on the already transformed augmented matrix.
+This came from an overbroad textual replacement during audit extension.
+Fix the dispatch to OP==8 ONLY, rerun restricted/unrestricted inv_pivoted,
+and correct conclusions/records before treating it as a standalone defect.
+The current pivoted-inverse stall evidence is for that two-call composition.
+Other operation results are unaffected by this particular defect, but audit
+all generated harness edits carefully before promoting them to regressions.
+
+Experiment in progress at pause (temporary copy only):
+/tmp/glass-fence-experiment is a git-archive copy of baseline GLASS.
+BlockBarrier::sync has empty asm volatile memory clobbers before/after sync.
+The same clobbers surround raw __syncthreads in softmax/inv/ldlt/syev there.
+Purpose: test compiler-visible memory handoffs without adding hardware barriers.
+This is a diagnostic experiment, NOT an adopted or validated library fix.
+Compile command:
+ nvcc -O3 -std=c++17 -arch=sm_120 -I/tmp/glass-fence-experiment \
+ /home/plancher/Desktop/GLASS/docs/audits/2026-10-28-restrict/extended.cu \
+ -o /tmp/glass-fence-experiment/audit
+Log: /tmp/glass-fence-experiment/compile.log
+Unified exec session 86596 was still compiling at last poll. No GPU run begun.
+Its current input still has the OP23 dispatch defect; correct before interpreting.
+
+Next steps on resume:
+1. Inspect compile status; repair/audit exploratory harness and isolated repros.
+2. Build reliable per-case permanent regression runner: numerical failures must
+   exit nonzero, fatal cases isolated/time-bounded, expected cases accounted for.
+3. Diagnose memory handoffs, independent restrict qualifiers, standalone CUDA
+   reduction; compare generated code as needed. Test candidate repairs and
+   measure performance before broad adoption.
+4. Integrate tests/cache/shards/obligations, all planned correctness checks,
+   downstream checks and fresh signed full receipt before pushing source edits.
+5. Prepare 0.5.0 release separately; don't conflate Git push with publication.
+
+Tool access: GPU nvidia-smi and CUDA execution require sandbox escalation;
+inside sandbox nvidia-smi misleadingly says driver unavailable. Outside sandbox
+RTX 5090 sm_120, nvcc 13.2.86, driver 615.71.09. gh network also escalated.
+No agents spawned; delegation not explicitly authorized.
+
+Instructions already consulted: GLASS CLAUDE.md, debugging guide (large combined
+read truncated; reread needed sections), testing_oracles.rst, namespace guide,
+bench/TUNING.md (combined output may have truncated). No AGENTS.md in GLASS.
+For edits, honor arbitrary legal thread counts, barrier policy for shared bodies,
+thread-only no-op sync, trailing sync ownership, generated manifest last,
+no handwritten generated tuning data. Required docs strict build and tests.