Repository navigation

GPU boundary, recurrence-derived prefill/decode, site geometry, and mapal fmt - #1

Merged

LessComplexity merged 324 commits into

mainfrom

site-geometry-latest

Nov 1, 2026

Merged

GPU boundary, recurrence-derived prefill/decode, site geometry, and mapal fmt#1
LessComplexity merged 324 commits into
mainfrom
site-geometry-latest

This pull request is big! We're only showing the most recent 250 commits

Commits on Oct 21, 2026
Commits on Oct 22, 2026
- test(gpu): record mixed small-M results
  LessComplexity
  committed
  bc7af76
- docs(gpu): preserve parallel validation prep
  LessComplexity
  committed
  29ad618
- chore(perf): organize reports and raw evidence
  LessComplexity
  committed
  dece72f
- perf: record competitive library comparisons
  LessComplexity
  committed
  d80d8f5
- fix(gpu): select policy for emitted arithmetic
  LessComplexity
  committed
  50a57e5
- docs: record reviewed observer repair and handoff
  LessComplexity
  committed
  293237a
- docs: record executed library kernel evidence
  LessComplexity
  committed
  4e1102c
- feat(cli): mapal build|run|dump-ir|test
  
  LessComplexity
  and
  claude
  committed
  8f1ae02
- bench(gpu): runtime measurement harness — request clock, retention loops, enqueue trace
  
  LessComplexity
  and
  claude
  committed
  b29061f
- feat(gpu): cell-fill and asyncstage rungs
  
  LessComplexity
  and
  claude
  committed
  11c56ef
- feat(gpu): ordered tile geometry rungs from shape + device facts
  
  LessComplexity
  and
  claude
  committed
  d380843
- feat(gpu): opt-in reassoc face, deterministic split-K
  
  LessComplexity
  and
  claude
  committed
  6c3162c
- bench(gpu): S62 runtime raw results, RTX 4090 and RTX 5090 session 1
  
  LessComplexity
  and
  claude
  committed
  56f90f3
- bench(gpu): S62 runtime raw results, RTX 4090 and RTX 5090 confirm sessions
  
  LessComplexity
  and
  claude
  committed
  ff052fc
- bench(gpu): submission stage 2, generated vs direct launch enqueue
  
  LessComplexity
  and
  claude
  committed
  ba7d850
- bench(gpu): S62 runtime raw results, stage-2 submission rows and RTX 3090 confirm half A
  
  LessComplexity
  and
  claude
  committed
  9bb2f07
- bench(gpu): S62 runtime raw results, H100 PCIe session 1, H100 SXM confirm part 1, H100 stage 2
  
  LessComplexity
  and
  claude
  committed
  fbf2c75
- bench(gpu): S62 runtime raw results cut by the 19:25 ssh drop
  
  LessComplexity
  and
  claude
  committed
  79e5869
Commits on Oct 23, 2026
- bench(gpu): S62 runtime raw results, the 19:27 completion sessions cut by the session restart
  
  LessComplexity
  and
  claude
  committed
  dad4a56
- feat(gpu): host CPU profile in GPU builds, per-site placement
  
  LessComplexity
  and
  claude
  committed
  daeef18
- perf(gpu): asyncstage measured on four parts
  
  LessComplexity
  and
  claude
  committed
  be89bc7
- perf(gpu): record S62 hostopt/place runs on four parts
  
  LessComplexity
  and
  claude
  committed
  f8d7272
- bench(gpu): S62 runtime raw results, the 20:00 completion sessions and merged session files
  
  LessComplexity
  and
  claude
  committed
  e5d6337
- bench(gpu): S62 runtime raw results, retained vs regenerated on one box per cut session
  
  LessComplexity
  and
  claude
  committed
  2733746
- perf(gpu): S62 runtime verdict — allocation latency, submission, retention
  
  LessComplexity
  and
  claude
  committed
  2b25583
- fix(gpu): gate fill-rows and panel-regs on geometry facts
  
  LessComplexity
  and
  claude
  committed
  cc49cae
- perf(gpu): ship the S62 geometry rungs per part from the v2 A/B
  
  LessComplexity
  and
  claude
  committed
  c862247
- test(gpu): 4090 64^3 snapshots back to the S53 launch; cite traced probe ms
  
  LessComplexity
  and
  claude
  committed
  5dff8df
- perf(gpu): S62 geometry raw results v1-v4 on 5090/4090/H100 PCIe/3090
  
  LessComplexity
  and
  claude
  committed
  580f672
- Merge branch 's62/geometry' into s62-integrated
  LessComplexity
  committed
  40f858c
- Merge branch 's62/asyncstage' into s62-integrated
  
  LessComplexity
  and
  claude
  committed
  ee5770d
- Merge branch 's62/splitk' into s62-integrated
  
  LessComplexity
  and
  claude
  committed
  c0eaa84
- Merge branch 's62/cpugpu' into s62-integrated
  
  LessComplexity
  and
  claude
  committed
  b94177c
- Merge branch 's62/runtime' into s62-integrated
  
  LessComplexity
  and
  claude
  committed
  52c57e0
- Merge branch 's62/cli' into s62-integrated
  
  LessComplexity
  and
  claude
  committed
  0601a93
- fix(gpu): cell-fill split partials land in their workspace slab
  
  LessComplexity
  and
  claude
  committed
  9a9cc20
- perf(gpu): S62 integration measured on four parts, per-part rung rows
  
  LessComplexity
  and
  claude
  committed
  6a0e3a0
- perf(gpu): RTX 3090 executed-code and usage evidence, S62 integrated
  
  LessComplexity
  and
  claude
  committed
  b762272
- perf(gpu): RTX 5090 executed-code and usage evidence, S62 integrated
  
  LessComplexity
  and
  claude
  committed
  999b89e
- perf(gpu): RTX 4090 executed-code and usage evidence, S62 integrated
  
  LessComplexity
  and
  claude
  committed
  98a782a
- perf(gpu): H100 PCIe executed-code and usage evidence, S62 integrated
  
  LessComplexity
  and
  claude
  committed
  569b5e3
- docs: reconcile S62 integration, gap map and handoff
  
  LessComplexity
  and
  claude
  committed
  3ec494e
- feat(float): the default float face — fma, order free, bounded (ADR-0040)
  
  LessComplexity
  and
  claude
  committed
  5c06a78
- perf(gpu): S63 default-face rungs measured on four parts, OFF rows
  
  LessComplexity
  and
  claude
  committed
  c0103c5
- wip(evidence): S64 intent registry, checker, plan/host evidence lines
  
  LessComplexity
  and
  claude
  committed
  068852d
- feat(evidence): checker fixes from the first static pass, box arm wiring
  
  LessComplexity
  and
  claude
  committed
  b2645e4
- perf(evidence): static SASS / ptxas / remarks for the suite and the re-judge cells, four parts
  
  LessComplexity
  and
  claude
  committed
  b88e209
- perf(evidence): RTX 4090 box run — suite, cells, MLPs, re-judge arms, torch, JIT == ptxas
  
  LessComplexity
  and
  claude
  committed
  59a90b3
- perf(evidence): RTX 5090 box run — suite, cells, MLPs, re-judge arms, torch; JIT != ptxas 12.8
  
  LessComplexity
  and
  claude
  committed
  d90eff1
- perf(evidence): H100 PCIe box run — suite, cells, MLPs, re-judge arms, torch, JIT == ptxas
  
  LessComplexity
  and
  claude
  committed
  8d758a9
- perf(evidence): RTX 3090 box run — suite, cells, MLPs, re-judge arms, torch, JIT == ptxas
  
  LessComplexity
  and
  claude
  committed
  a1944a8
- perf(gpu): warp-fill class split on H100 / 3090 (Under -> Partial), S64 row evidence
  
  LessComplexity
  and
  claude
  committed
  7b6a889
- docs(evidence): S64 end check, report and handoff; 3090 warp-fill split not taken
  
  LessComplexity
  and
  claude
  committed
  38789b4
- wip(s65): loop state by reference via byref.rs (reference; to be folded into the existing allocators)
  
  LessComplexity
  and
  claude
  committed
  de7226f
Commits on Oct 24, 2026
- feat(s65): rule R in the existing storage layer; byref.rs deleted
  
  LessComplexity
  and
  claude
  committed
  c66c50e
- feat(s65): compact host allocator; device leg of the property test
  
  LessComplexity
  and
  claude
  committed
  a7dfff7
- bench(s65): static evidence on four archs, RTX 4090 box results, movement and arm reports
  
  LessComplexity
  and
  claude
  committed
  2c5c0e8
- bench(s65): RTX 5090 and RTX 3090 box results; movement table grouped by arch
  
  LessComplexity
  and
  claude
  committed
  183032e
- bench(s65): H100 PCIe box results; instances after the run
  
  LessComplexity
  and
  claude
  committed
  a972f76
- perf(gpu): retain-resident ON on the 4090 and 5090 (S65 rows removed)
  
  LessComplexity
  and
  claude
  committed
  9ca6c6b
- fix(evidence): the null arm's runtime movement is predicted by the shipped stem's facts
  
  LessComplexity
  and
  claude
  committed
  78742f4
- docs(s65): evidence README section and next-session handoff
  
  LessComplexity
  and
  claude
  committed
  c8aae2f
- bench: commit the generated algorithm programs under benches/algorithms
  
  LessComplexity
  and
  claude
  committed
  249f199
- wip(natural-forms): first natural-form generators, natural_forms test, elem_plan/write_delta edits (interrupted for the split)
  
  LessComplexity
  and
  claude
  committed
  8b499e6
- ir: footprint_plan — every consumer's read edges as one query
  
  LessComplexity
  and
  claude
  committed
  30f661f
- natural-forms: piecewise map law over a stepped cell set; column write in place
  
  LessComplexity
  and
  claude
  committed
  52cb157
- llvm: generic fusion rule (fuse), window rung, strided 2-D sites
  
  LessComplexity
  and
  claude
  committed
  c94929a
- bench(natural-forms): idiomatic decode and softmax generators, regenerated snapshot
  
  LessComplexity
  and
  claude
  committed
  6b04916
- ir/llvm: window row folds; bench: cross-algorithm chain workloads
  
  LessComplexity
  and
  claude
  committed
  802fb15
- llvm: window row folds behind the window rung
  
  LessComplexity
  and
  claude
  committed
  e3a5eca
- test: default-face on == off in the chain property test; rustfmt the new files
  
  LessComplexity
  and
  claude
  committed
  9c3b969
- bench: the chain workloads in the benches/algorithms snapshot (drift-free)
  
  LessComplexity
  and
  claude
  committed
  467c865
- natural-forms: twin property test over 256 programs; range-map rung; dead trip work in the loop
  
  LessComplexity
  and
  claude
  committed
  4186713
- test(natural-forms): the twin sweep's reach floor at its default size only
  
  LessComplexity
  and
  claude
  committed
  e203c5d
- natural-forms: a delta wider than its array is the whole upload
  
  LessComplexity
  and
  claude
  committed
  89d5262
- testgen(natural-forms): twin range writes may run into the next row
  
  LessComplexity
  and
  claude
  committed
  ef7a5f6
- test(natural-forms): the range-map edge programs at 16-wide rows (the debug oracle)
  
  LessComplexity
  and
  claude
  committed
  26f2f68
- natural-forms: the emitter asks each map's piecewise law once
  
  LessComplexity
  and
  claude
  committed
  14779ad
- merge natural-forms into generic-fusion
  
  LessComplexity
  and
  claude
  committed
  bade438
- fix(fuse): a unit kernel carries its cells' nested bodies
  
  LessComplexity
  and
  claude
  committed
  8d1f66b
- bench: regenerate the snapshot — attention's exp in natural form
  
  LessComplexity
  and
  claude
  committed
  2924d14
- evidence: window and range-map intent rows
  
  LessComplexity
  and
  claude
  committed
  a55ead9
- fix(by-ref): an in-place range map owns its bytes, as an Update does
  
  LessComplexity
  and
  claude
  committed
  fbdec8b
- evidence(fusion-natural-forms): static evidence of the integrated build
  
  LessComplexity
  and
  claude
  committed
  8dac5b5
- evidence(fusion-natural-forms): CPU-only builds identical to both parts'
  
  LessComplexity
  and
  claude
  committed
  12727a4
- test(natural-forms): a fold over a range stays a device cell, in a unit too
  
  LessComplexity
  and
  claude
  committed
  d2dd44c
- docs(evidence): fusion-natural-forms test results at d2dd44c
  
  LessComplexity
  and
  claude
  committed
  47a5048
- bench: library legs for the generic-fusion chains
  
  LessComplexity
  and
  claude
  committed
  dc0372e
- fix(bench): attention's p buffer was the operand p
  
  LessComplexity
  and
  claude
  committed
  265fe69
- evidence(fusion-natural-forms): H100 PCIe and RTX 5090 box runs
  
  LessComplexity
  and
  claude
  committed
  edf96e9
- evidence(fusion-natural-forms): attention re-take, H100 PCIe and RTX 5090
  
  LessComplexity
  and
  claude
  committed
  8350361
- evidence(fusion-natural-forms): RTX 4090 box run
  
  LessComplexity
  and
  claude
  committed
  b8796e8
- evidence(fusion-natural-forms): RTX 3090 box run
  
  LessComplexity
  and
  claude
  committed
  f007a6f
- evidence(fusion-natural-forms): RTX 5090 confirm box (Ryzen 9 7945HX host)
  
  LessComplexity
  and
  claude
  committed
  cfb1458
- evidence(fusion-natural-forms): the 5090 confirm box throttled mid-run
  
  LessComplexity
  and
  claude
  committed
  7400a00
- perf(gpu): fuse ships OFF on all four measured parts
  
  LessComplexity
  and
  claude
  committed
  cbe0a66
- evidence(fusion-natural-forms): RTX 3090 confirm box (EPYC 7B12, 250 W)
  
  LessComplexity
  and
  claude
  committed
  e2b0e68
- evidence(fusion-natural-forms): the slow 3090 confirm box to confirm-throttled/
  
  LessComplexity
  and
  claude
  committed
  678caee
- evidence(fusion-natural-forms): RTX 4090 confirm box (Ryzen 9 5950X, 375 W)
  
  LessComplexity
  and
  claude
  committed
  4898e15
- evidence(fusion-natural-forms): RTX 5090 confirm box (Ryzen 9 7950X, 575 W)
  
  LessComplexity
  and
  claude
  committed
  d071cbc
- evidence(fusion-natural-forms): RTX 3090 confirm box on S65's own host (EPYC 7702, 350 W)
  
  LessComplexity
  and
  claude
  committed
  c10c1e4
- evidence(fusion-natural-forms): H100 PCIe confirm box on S65's own host (EPYC 7713)
  
  LessComplexity
  and
  claude
  committed
  a7fa76c
- perf(gpu): the fuse rows cite the second hosts; sweep tables
  
  LessComplexity
  and
  claude
  committed
  bd546e9
- docs: the fusion-natural-forms sweep — evidence README, next-session, session log
  
  LessComplexity
  and
  claude
  committed
  9a3477a
- perf(ir): facts live on the graph
  
  LessComplexity
  and
  claude
  committed
  285f61a
- perf(gpu): a kernel no launch site loads is not emitted
  
  LessComplexity
  and
  claude
  committed
  349767a
- bench: causal attention decode, prefill+decode, multi-channel conv
  
  LessComplexity
  and
  claude
  committed
  1a881e8
- bench(gpu): facts-on-graph sweep on 4 archs, default face
  
  LessComplexity
  and
  claude
  committed
  9c1e63a
- fix(bench): default face gated by its bound, never bit for bit
  
  LessComplexity
  and
  claude
  committed
  072f726
- refactor(gpu): rung fuse is flow
  
  LessComplexity
  and
  claude
  committed
  531eb9c
- perf(gpu): a flow unit's row is a block of lanes
  
  LessComplexity
  and
  claude
  committed
  87d9739
- test(gpu): default-face flow rows are lanes with tree folds
  
  LessComplexity
  and
  claude
  committed
  c470546
- fix(gpu): a retained output's move intent is its steady trip
  
  LessComplexity
  and
  claude
  committed
  437dcd5
- bench(gpu): torch legs for pj/twin/epi, a stem without a torch leg is a row not an exit; SWEEP_PROF nsys/SASS section
  
  LessComplexity
  and
  claude
  committed
  81857cc
Commits on Oct 25, 2026
- bench(gpu): flow sweep on 4 archs (flow ON/OFF, retain-resident h100 re-judge, nsys per-kernel of every workload)
  
  LessComplexity
  and
  claude
  committed
  5eaad72
- docs(evidence): why-slow — per losing workload x arch gap split, causes, fixes ranked by measured ms
  
  LessComplexity
  and
  claude
  committed
  04a5cf3
- perf(gpu): flow ships ON on all four parts
  
  LessComplexity
  and
  claude
  committed
  7a400a4
- docs: handoff to cpu-gpu-boundary; commit the agent rules (CLAUDE.md, AGENTS.md)
  
  LessComplexity
  and
  claude
  committed
  adc4409
- docs: drift — renamed split_count and in_place_source rows
  
  LessComplexity
  and
  claude
  committed
  5c8093a
- perf(gpu-rt): kernels, arena and uploads leave the timed path
  
  LessComplexity
  and
  claude
  committed
  64de5f0
- perf(gpu): uniform and nest-cell placement, tree-gate
  
  LessComplexity
  and
  claude
  committed
  c8a49d6
- merge cpu-gpu-boundary-runtime into placement
  
  LessComplexity
  and
  claude
  committed
  317bdf1
- fix(cpu-gpu-boundary): review findings of both builders
  
  LessComplexity
  and
  claude
  committed
  8a4142f
- fix(evidence): entry-arena OFF check reads ON-only text (arena base load)
  
  LessComplexity
  and
  claude
  committed
  e8c2050
- fix(gpu-bench): atomic split launch line; entry-arena ON applies where an arena exists; vast LABEL
  
  LessComplexity
  and
  claude
  committed
  3190d8c
- fix(gpu-bench): differential ignores the clock line; TF32 atomic split-K within 1e-4; no macOS oracles uploaded
  
  LessComplexity
  and
  claude
  committed
  16974f9
- fix(evidence): a slot the entry arena seeds never allocates (runtime alloc prediction)
  
  LessComplexity
  and
  claude
  committed
  8d6b215
- results(cpu-gpu-boundary): rtx-4090 smoke — values, ABBA arms, evidence
  
  LessComplexity
  and
  claude
  committed
  1169548
- bench(gpu): integrated_bundle drops a rung arm by emission before compiling it
  
  LessComplexity
  and
  claude
  committed
  15f52cf
- bench(gpu): cpu-gpu-boundary sweep on 4 archs (value gates, ON/OFF of every new rung, retain-resident O2, nsys per-kernel)
  
  LessComplexity
  and
  claude
  committed
  73739a7
- docs(evidence): cpu-gpu-boundary end check, rows, refreshed gap table
  
  LessComplexity
  and
  claude
  committed
  9bb981c
- docs(readme): GPU workload ratios on 4 archs
  
  LessComplexity
  and
  claude
  committed
  8e41f94
- perf(gpu): rung rows from the cpu-gpu-boundary sweep
  
  LessComplexity
  and
  claude
  committed
  8ef1de8
- docs: cpu-gpu-boundary session log and handoff; fix swapped ON/OFF in the H100 retain row
  
  LessComplexity
  and
  claude
  committed
  f7925b4
- docs(adr): ADR-0041 candidate — compute precision as a permission on compute nodes
  
  LessComplexity
  and
  claude
  committed
  ab82876
- docs: next-session item 1 — prefill/decode emerge from the loop recurrence
  
  LessComplexity
  and
  claude
  committed
  8e876a9
- docs(adr): ADR-0041 accepted for implementation
  
  LessComplexity
  and
  claude
  committed
  0736051
- docs(adr): ADR-0041 surface — a type on the compute step (fold: tf32 | f32); output type sets as later
  
  LessComplexity
  and
  claude
  committed
  71bf770
- docs(adr): ADR-0041 — ': compute' on steps, blocks and functions
  
  LessComplexity
  and
  claude
  committed
  7a7f500
- docs(adr): ADR-0041 — function compute type follows the output (fn f(a) -> T: compute)
  
  LessComplexity
  and
  claude
  committed
  29f0a54
- test(llvm): bound the differential harness to the core budget; time-box emulated runs
  
  LessComplexity
  and
  claude
  committed
  0d20476
- feat(ir,gpu): mixed-radix index fact; convmc as an implicit-GEMM tile
  
  LessComplexity
  and
  claude
  committed
  7c25436
- feat: ADR-0041 compute precision as `: compute` on bodies; tf32-tile rung
  
  LessComplexity
  and
  claude
  committed
  1390658
- build: optimize mapal-interp in dev/test builds (the oracle), checks kept
  
  LessComplexity
  and
  claude
  committed
  70441f0
- docs: next-session — test-speed item with measurements
  
  LessComplexity
  and
  claude
  committed
  dbfbfd5
- perf(gpu): prefill and decode emerge from the loop recurrence
  
  LessComplexity
  and
  claude
  committed
  85dba44
- perf(gpu): nest cells and row folds fold over a warp's lanes
  
  LessComplexity
  and
  claude
  committed
  4434029
- Merge branch 'recurrence-precision-tiles' into recurrence-precision
  LessComplexity
  committed
  bd5f7f2
- Merge branch 'cpu-gpu-boundary-latest' into recurrence-precision
  LessComplexity
  committed
  87e1295
- merge: resolve the mapal-ir re-export conflicts (TripSplit with PrecisionPlan/Radix*)
  
  LessComplexity
  and
  claude
  committed
  6ab43e3
- fix(review): radix sites — one output, tensor branch, admission arms; tf32 test face
  
  LessComplexity
  and
  claude
  committed
  578e06d
- fix(review): trip-batch keeps traps, batches invariants; range-cell guard reads
  
  LessComplexity
  and
  claude
  committed
  f619fe2
- docs(ir): the lanes test row names lanes_on_unseen_programs (drift)
  
  LessComplexity
  and
  claude
  committed
  674861f
Commits on Oct 26, 2026
- fix(gpu): tensor cores and fused radix steps only where the order is free
  
  LessComplexity
  and
  claude
  committed
  436836a
- fix(gpu): a +0.0 fill carry init is zeroed on the device
  
  LessComplexity
  and
  claude
  committed
  e47f685
- test(review): recurrence generator shapes; lanes gated on the interpreter
  
  LessComplexity
  and
  claude
  committed
  c8215b5
- fix(gates): M through relu and bias chains; trip-batch in the movement prediction
  
  LessComplexity
  and
  claude
  committed
  c99ac7f
- results(recurrence-precision): rtx-4090 smoke — values, ABBA arms, evidence
  
  LessComplexity
  and
  claude
  committed
  c6cd2d1
- bench(torch): TF32-on legs (matmul + cuDNN allow_tf32) for tf32 / conv / mlp programs beside the FP32 ones
  
  LessComplexity
  and
  claude
  committed
  7d97152
- results(recurrence-precision): sweep step 1 — emission diff vs cpu-gpu-boundary-latest, changed programs and differing OFF arms
  
  LessComplexity
  and
  claude
  committed
  006cf8c
- bench(torch): torch legs for pt_gen (prefill math), the _tf32 programs and the mlp / mlp2 chains
  
  LessComplexity
  and
  claude
  committed
  87bc83c
- results(recurrence-precision): narrowed sweep on 4 archs — changed programs only, ON/OFF per new rung, torch TF32 legs, nsys
  
  LessComplexity
  and
  claude
  committed
  784e69d
- docs(recurrence-precision): end check, gap table, what moved, open
  
  LessComplexity
  and
  claude
  committed
  a06c512
- docs(readme): GPU results after recurrence-precision
  
  LessComplexity
  and
  claude
  committed
  d2a9ab2
- rungs(rtx-4090): fold-lanes ships OFF
  
  LessComplexity
  and
  claude
  committed
  a2f817d
- docs: recurrence-precision session log and handoff
  
  LessComplexity
  and
  claude
  committed
  1950c54
- docs(adr): ADR-0034 accepted for implementation (sprint site-geometry)
  
  LessComplexity
  and
  claude
  committed
  1613a13
- geom(B1): per-target geometry table with --geom overrides
  
  LessComplexity
  and
  claude
  committed
  874944f
- bench(changed): emission diff per arch x face vs a base ref, arm tags, result cache
  
  LessComplexity
  and
  claude
  committed
  be93a32
- bench(bundle): changed-only bundles and a preflight before any rental
  
  LessComplexity
  and
  claude
  committed
  22b8170
- bench(search): ADR-0034 D2 search driver over benches/gpu/knobs.json
  
  LessComplexity
  and
  claude
  committed
  f30cc05
- fix(preflight): value_gate.py only where a probe answers to D2's bound
  
  LessComplexity
  and
  claude
  committed
  107017b
- test(device-loops): run the programs over the job slots; touched-files helper
  
  LessComplexity
  and
  claude
  committed
  785f2ad
- geom(B2): fill-wave and row-split — one fill decision for tile sites
  
  LessComplexity
  and
  claude
  committed
  f9c4790
- geom(R1): fit-first — reshape a tall thread grid before shrinking it
  
  LessComplexity
  and
  claude
  committed
  4eddb74
- geom(B4): GEMM tile candidates as table entries
  
  LessComplexity
  and
  claude
  committed
  0bffbd2
- geom(B3): tf32-radix — the mma.sync tile for a tf32 radix site
  
  LessComplexity
  and
  claude
  committed
  1f932b9
- Merge branch 'site-geometry-measure' into site-geometry
  LessComplexity
  committed
  bf6f103
- fix(geom): review findings on site-geometry-sites
  
  LessComplexity
  and
  claude
  committed
  f504761
- fix(measure): review findings on site-geometry-measure
  
  LessComplexity
  and
  claude
  committed
  ff750a8
- fix(changed): runtime stamp over what a row's time runs through (runtime, timing harnesses)
  
  LessComplexity
  and
  claude
  committed
  e275f17
- fix(gen_tiles): K through a map's output (a fold over a computed array)
  
  LessComplexity
  and
  claude
  committed
  434c571
- fix(evidence): atomic adds with memory-order qualifiers (newer llc); smoke plumbing (TORCH passthrough, --rungs on-arms)
  
  LessComplexity
  and
  claude
  committed
  8e9b432
- fix(gpu): an epilogue group's ordered launch keeps its own tile beside a fill-wave split
  
  LessComplexity
  and
  claude
  committed
  6a37cd9
- fix(bundle): cells keep their reassociation faces under SUITE_ARMS=0 (box_run's split-aware cell section keys on _r)
  
  LessComplexity
  and
  claude
  committed
  e07a812
- bench(site-geometry): rtx-4090 smoke — value gates, ON/OFF of the new rungs
  
  LessComplexity
  and
  claude
  committed
  a9c16a9
- wip(site-geometry): sweep evidence collected before a checkpoint (laptop closed) — 4 archs, pass2 results, evidence, ncu/traces; summary not written yet
  
  LessComplexity
  and
  claude
  committed
  3526d41
- docs(rules): a mixed ON/OFF result is a missing fact, not an OFF row
  
  LessComplexity
  and
  claude
  committed
  f794649
- bench(site-geometry): sweep h100 value-gated re-run, search picks, report
  LessComplexity
  committed
  9d550f5
- bench(site-geometry): h100 search run, picks with shipped values, report
  LessComplexity
  committed
  037b855
- perf(gpu-rungs): site-geometry judge — tf32-radix ON on 5090/3090, row-split and fit-first stopgaps
  
  LessComplexity
  and
  claude
  committed
  9583482
- fix(box_run): the cuBLAS timing section's _exact face reads the _x module
  
  LessComplexity
  and
  claude
  committed
  0035aad
- perf(gpu-rungs): keep tf32-radix OFF on the 5090 — the fitted FFMA tile is faster
  
  LessComplexity
  and
  claude
  committed
  1f1869e
- docs(evidence): site-geometry end check, gap table, missing facts
  
  LessComplexity
  and
  claude
  committed
  750c97f
- docs(readme): GPU workloads table from the site-geometry sweep
  
  LessComplexity
  and
  claude
  committed
  a25a029
- merge: recurrence-precision-latest (CLAUDE.md mixed-result rule) into site-geometry-latest
  
  LessComplexity
  and
  claude
  committed
  ba4f295
- results(cache): measurement cache from the site-geometry sweep (keyed by emission hash; lets the next sweep carry unchanged rows)
  
  LessComplexity
  and
  claude
  committed
  991e59a
- docs: site-geometry session log and handoff
  
  LessComplexity
  and
  claude
  committed
  e3d4710
- docs: next-session — fix the handoff text (site-geometry standing and missing facts)
  
  LessComplexity
  and
  claude
  committed
  1dbe09e
Commits on Oct 29, 2026
- docs: summary.md and handoff for the pull request to main
  
  sapiritur
  and
  claude
  committed
  3fc0750
- feat(fmt): mapal fmt source formatter (syntax, CLI, editors) and docs
  
  sapiritur
  and
  claude
  committed
  d4e34c6
- evidence: raw run output moves to a GitHub release; manifest, index, archive script, ignore rules
  
  sapiritur
  and
  claude
  committed
  3966475
Commits on Nov 1, 2026
- style: cargo fmt --all (CI formatting step)
  
  sapiritur
  and
  claude
  committed
  7df88ff
- test(retain_delta): contract oracle fuses on x86-64 (-mfma)
  
  sapiritur
  and
  claude
  committed
  a37c984

Commits

Commits on Oct 21, 2026

Commits on Oct 22, 2026

Commits on Oct 23, 2026

Commits on Oct 24, 2026