Add normative PTO v0.2 SYNCALL<core_scope> / FENCE.D.CORE4 definition #360

Activity

1. ckwllawliet commented on Oct 8, 2026

   

   ckwllawliet

   on Oct 8, 2026

   Last edited by ckwllawliet

   Collaborator

   More actions

   Accepted design r2 — SYNC.CORE ordinary Scalar SYS phase barrier

   This records the repository owner's accepted design review of the earlier issue proposal. It supersedes the proposed FENCE.D.CORE4/alias/wildcard form and target-profile gate. Baseline: e182c9b. The decision record is distinct from the implementation and validation status recorded below.

   Instruction: ordinary Scalar SYS SYNC.CORE, no operands, exact L32 MATCH 0x0330A02B, MASK 0xFFFFFFFF. Legal V1 envelope is exactly BSTART.SYS FALL / SYNC.CORE / BSTOP, without additional header modifiers/body operations; validate the complete envelope before arrival, preserving existing fetch/illegal/Block fault classes. Preserve FENCE.D/FENCE.I encoding and semantics. No FENCE.D.CORE4, dsb.core4, variable pred/succ, runtime mask/count/ID, new selectable architecture/profile or capability gate. V5 SYNCALL<core_scope> lowering remains downstream mapping; scope-less SYNCALL() is unchanged.

   Execution and synchronization: four PEs share code and initial entry but independently maintain subsequent PC/Block/temporary/exception state. All PE0–PE3 of one Core and kernel invocation must reach the same static SYNC site in the same dynamic order. Each arrival is non-speculative, irrevocable and counted once. Divergent participation has no progress guarantee and never shrinks membership or falsely releases. Completion requires all arrivals plus all earlier relevant GM operations reaching required completion/visibility, including not-yet-issued TLSU work and producer dependencies. It creates real participant-pair GM RW->RW synchronization, not four local fence events or an already-issued queue drain. No physical DRAM flush, inter-Core rendezvous, Tile publication, device/IC synchronization or implicit GMOV/matmul synchronization. Preserve existing coherence, atomicity, shareability and mixed-size limitations.

   Accepted recovery/lifecycle details:

   1. Older work which can still cause recovery before SYNC prevents irrevocable arrival. Validated non-faulting but incomplete/unissued work can remain and prevent readiness. Fault/retry is never successful readiness.
   2. One active kernel per Core in V1; no kernel switching/migration. Launch establishes a new synchronization instance. Recoverable fault/interrupt retains it; abort/reset permanently discards it without rolling back committed GM writes.
   3. An arrived PE may handle an interrupt while its kernel context stays parked. Original arrival survives. Return waits or resumes once depending on completion; repeated generations preserve slow/interrupted peers' unconsumed releases. Handler SYNC reentry is rejected before collective changes.
   4. GM operations performed by a handler after arrival are outside this barrier and gain no collective completion/publication guarantee. Event provenance must prevent incidental event-index order from admitting them.
   5. Successful completion clears each PE's current LR/SC reservation, including one newly established in a handler. Trap return must not resurrect it. No IC epoch or unrelated SYS effects.
   6. Before a later kernel launch, previous work must complete or be safely canceled. Reversible work may be canceled; irrevocable work completes before launch. Old completion cannot update a new synchronization instance.

   Owner/decision plan: one new accepted ADR-SCALAR-0007 with interface_change=true for the new Scalar interface, secondary state/Block/memory ownership and protected ADR-GOV-0012 (no target-profile resurrection). Existing ADR-SCALAR-0006 retains local-fence ownership. New primary owners: asl/scalar/sys/SYNC.CORE.asl, asl/arch/state/{pe-execution-context,core-synchronization,kernel-lifecycle}.asl, asl/arch/memory-model/gm-obligations.asl; extend actual encoded Scalar/Block dispatch, program control/trap/fault paths and memory candidate checker. IDs: PTO-SYNC-CORE-DECISION-BINDING-001, PTO-ARCH-PE-EXECUTION-CONTEXT-001, PTO-ARCH-CORE-SYNCHRONIZATION-001, PTO-ARCH-KERNEL-LIFECYCLE-001, PTO-ARCH-GM-OBLIGATIONS-001, PTO-ARCH-MEMORY-MODEL-COLLECTIVE-SYNC-001. Current meaning will be owned by these ASL/NDF units; this comment/ADR are decision records.

   Required focused evidence: exact decode/handler and full overlap/near-neighbor rejection; legal/invalid/unfetchable complete envelopes; independent PE context and encoded arrival/wait/resume; 1/2/3 arrival nonrelease; delayed unissued producer/store and outstanding async GM; precise fault/retry and wrong-path/replay; independent cross-agent stale-read rejection including mixed Scalar/TLSU/atomic paths; loop generations/fast reuse/slow release; PE_MASK zero and Core/invocation isolation; interrupt/nested return/handler exclusion/reentry; current reservation invalidation; abort/quiescence/next launch and old completion rejection. Preserve protected existing instruction and memory behavior. Model array bounds are verification bounds, not hardware limits.

   Compatibility/release: additive encoding plus explicit closure of missing independent PE-context semantics; no external backend edits or ASLRef pin movement. V1 specification PR with release impact required, no V2 release-validation/tag/publication claim. Final signed candidate requires focused executable ASL and decoder evidence, idempotent generated projections, prepare-pr-selected lightweight gates and binary closure, main owner acceptance, and independent exact-input review. LLVM #116, SuperScalarModel #894 and ASL-MODEL compile/link adoption remain separately tracked downstream obligations. No remaining open user architecture decision from this design review; genuine discovered conflicts return to the main architecture owner for recorded reconciliation.

   Implementation-owner clarification (r2.1)

   The baseline TSTORE path is synchronous. To provide executable evidence for deferred earlier GM work, permit a minimal architecture submission/obligation/progress carrier and verification scheduler which defaults to immediate progress. It must capture actual encoded request/descriptor/source-generation and provenance, then progress through real producer/probe/store/completion/fault paths. It is not a body-local queue or ISA scheduling control; normal results and precise restart remain unchanged. All relevant GM classes participate, and earlier recovery risk prevents irrevocable arrival. A bounded inventory of modeled Core instances may prove isolation; its size is a verification bound and never changes the four mandatory PE identities per Core. Core selection preserves system memory/history and isolates local contexts. No target profile or additional public instruction is introduced.

   Interleaved recovery closure (r2.2)

   Independent PE execution exposes singleton-recovery assumptions: discard only the faulting Core/PE/request's uncommitted replay records, preserving surviving event/read-from identities and committed effects. Ordinary trap return must not reapply whole stale Shared or cooperative Local generation snapshots. Keep the existing Core carrier with per-PE generation entries and existing scoped writer/parent abort rules; restore only the current PE's restart carriers. Execution-domain allocation stays unique across the Core. Focused evidence must preserve a peer's later readiness/publication and interleaved read-from reference through another peer's recovery. This is closure of the existing fault/generation isolation contract, with no new generic Tile rendezvous.

   Actual irrevocable-work witness (r2.3)

   The verification scheduler may pause a validated non-faulting captured TSTORE after an actual committed beat, defaulting to complete immediate progress. Abort preserves its committed effects and captured remaining work; the closing old invocation drains the real remaining beats with original request provenance, without resuming its kernel. Pending launch starts only after quiescence; the production completion entry rejects stale old identities from updating a later invocation. Faulted partial requests retain existing precise fault/retry semantics and are never declared complete merely by abort. If old work remains neither complete nor safely cancelable, launch stays waiting without a new fault or progress promise. This provides executable evidence for accepted Q9, without a new ISA control or fault-cancellation rule.

   Irrevocable memory history (r2.4)

   Generation-slot reuse must not erase completed collective ordering. Retained event provenance owns durable completion/phase history, including original Core, invocation and kernel/handler identity; active-generation records are insufficient for this history. Only completed barriers advance it. Delayed requests retain their submission identity, and comparisons create no collective edges across invocations or from handler events. Full ring reuse, later launch and Core selection must preserve old-event relations and stale-read rejection. Reset follows existing trace-retention rules. This repairs the accepted irrevocable-completion rule without adding synchronization semantics.

   State registry closure (r2.4 implementation note)

   Retain the existing state scope vocabulary and same-file declaration ownership: Core/PE banking represents existing logical ownership, and invocation is lifetime/classification rather than an additional scope domain. Partition shared Core collective/history carriers from private PE release/order carriers; ACR trap slots depend on PE context. Preserve strict scope/owner/member canaries. Current draft counts follow the new form, while historical release manifest counts and ABI identities retain their original immutable inputs.

   Independent-review closure plan (r2.5)

   The first complete independent review found eight production integration defects; passing V1 checks alone did not establish correctness. The same executor will fix real completion versus submission, typed captured-request progress, live-context preservation and real handler origin, ordinary system-wide monitor invalidation, inactive-event relations and invocation record reclamation. For independently fetched Tile/TLSU instances, retain the original logical PE_MASK and cooperative shape/publication rules while effects belong to the executing eligible PE; use authoritative peer payload views where existing aggregate primitives require them. Standalone aggregate verification helpers retain their established batch meaning. No new rendezvous, body queue, profile or publication rule is introduced. Permanent encoded regressions and the same independent reviewer's exact-input closure are required before PR publication; no new ADR boundary is needed beyond ADR-SCALAR-0007.

   Core candidate representation (r2.6)

   Permit a minimal operation-owned Core Shared candidate identity to preserve existing whole-candidate semantics when PE execution is independent. Keep original packing, publication legality/commit points, singleton whole-parent behavior and immediate publication from already-ready peer views. Per-operation source/provenance prevents fast/slow transfers from mixing contributions. Each direct operation retains its current partial-failure owner; B.ASSEMBLE's distinct rules are not generalized to ordinary transfers. This adds no body queue, new rendezvous or software-visible option. Representative executable closure precedes widening the correction sweep.

   Shared destination legality correction (r2.7 supersedes the r2.6 example)

   The authoritative PTO-B-ASSEMBLE-SHARED-STANDALONE-001 requires B.ASSEMBLE for multi-PE Shared destinations; direct destinations are single-PE. Preserve that rule: singleton whole-parent publication remains immediate, while full-mask destination mapping uses the existing assembly generation/INIT/LAST/coverage/participant owner. A Core candidate identity is only auxiliary representation for legal operations, not a new direct multi-PE path. The invalid full-mask direct-destination representative is replaced by encoded illegal-admission rejection plus legal four-stream B.ASSEMBLE closure.

   Captured-progress validation scope (r2.8)

   The reference model admits captured deferred TSTORE progress only after every active address passes its write probe; progress retains the original operands, mask and ACR, and this model has no mapping-revocation transition. F3 closure must therefore include actual handler/cross-context progress, real admission/retry fault tests and an independently checked reachability argument. An unreachable late-access-fault path will be recorded as such, not fabricated or claimed as executed coverage. No new public fault/profile semantics are introduced.

   Shared-operation lifecycle ownership (r2.9)

   The existing Shared-generation state owner will provide Core+invocation-scoped cleanup of actual-PE operation bookkeeping, called by kernel abort and successful end. Open candidates and completed/unconsumed release metadata are discarded while published Shared payload/descriptor, committed effects and operation ordinal remain; other Core/invocation and aggregate records are protected. Encoded contribution→abort/end→fresh-invocation evidence is required. This closes the accepted lifecycle integration within ADR-SCALAR-0007.

   Actual Shared contribution squash (r2.10)

   The existing execution-domain squash event will call a Shared-generation-owned identity hook. A matching incomplete actual-PE candidate is discarded using the existing parent-abort rule; published state, committed effects, ordinal allocator and other identities are preserved. Squashed tokens cannot re-contribute; real replay supplies a fresh request token. Published generations keep their irrevocable result and outstanding releases. Executable contribution→squash→fresh replay and protection witnesses are required; no partial rollback or new generic replay architecture is added.

   Replay validation through existing recovery (r2.12)

   Squash invalidates request data/identity without selecting a branch PC. The executable replay witness uses real trap save and existing architectural context-register recovery for a handler-chosen inactive-block restart; encoded BSTART then obtains a fresh token. Old restored canceled bindings cannot contribute again. No private-state test reset or new generic replay architecture is introduced.

   Complete implementation review correction batch (r2.12)

   The independent full review of signed f969ca9 returned changes_requested, F9–F14. The same executor will repair invocation-stale trap recovery, foreign-Core Shared cleanup, distinct dynamic TSTORE identity and first-successful-validation ACR capture, published-request completion retry, nonselected legal assembly, and the static AVS ID length. Seven real production probes reproduce the semantic failures; passing focused tests do not override them. These close the existing accepted contract without another public architecture boundary. The immutable-footprint late-fault argument is conditional and is not accepted for the rejected reuse path; corrected identity/admission guards will be independently verified. Same ADR/owners, complete follow-up review and exact-input PR gates remain required before publication. No release/downstream adoption claim.

   Existing TLOAD/Shared readiness ownership clarified within r2.12

   Direct affected owner: asl/tile/memory-and-data-movement/regular/TLOAD.asl, PTO-TLOAD-MEMORY-001 / PTO-TILE-TLOAD, now included in the same ADR inventory. Implementation corrections must preserve singleton whole-valid-parent GM reads, including B.ASSEMBLE. For multi-PE requests, each real PE keeps its own GM source stripe/private base/stride; the explicit B.ASSEMBLE offset is the destination range. Legal permuted offsets must work. The existing Shared-generation owner must check real data readiness independently of declared CELL coverage, preserving non-valid physical-tail rules and the prior published parent while valid payload is missing. No implicit offset-equals-PE restriction, new fault, profile or ADR boundary.

   The final batch includes encoded singleton TLOAD success, legal two-participant full-coverage/missing-data park, and all-four sparse singleton TMOV with undefined nonparticipant inputs. An earlier singleton quarter-read fixture asserted the wrong result and is retained only as defect evidence; a separate PE1 diagnostic was a Local/Shared binder-mask fixture mismatch. Working-source results do not certify the final signed candidate; complete independent review and exact-input gates remain pending.

   24. Complete round3 correction batch (r2.13; existing-owner repair)

   The same independent GPT-6.1 Sol Medium reviewer completed every planned area of fb4f15d and returned changes_requested, compatibility breaking on that rejected input, with consolidated F15–F17. The complete report, exact receipt, source/fixture/model identities, six probes and prior-model aggregate comparison are retained under candidate build/issue360-review-round3; snapshot the final report/receipt before replacing them. §§1–23 remain binding. Main accepts the complete batch as in-contract repairs; no new public architecture decision or ADR is allocated. Same sole executor and same independent reviewer continue.

   F15: exact already-admitted/published TMOV completion must retire without re-reading live source allocation/capacity/descriptor/payload, as required by F12. Retain applicable immutable wire/control/identity checks, structurally invalid/unrelated-token rejection and full source legality at first admission. Add real encoded handler source shrink 512B→256B and ReleaseTile, plus same-capacity type/shape boundaries: original stop consumes once, advances PC once, leaves parent/ordinal/peer releases stable and emits no repeated effects/events. Do not globally disable source validation.

   F16: preserve physical destination coordinates and all real valid elements across non-valid row/column tails. Undefined non-valid physical padding must not truncate later valid rows, be compacted into other coordinates, require fabricated values, or be confused with genuinely missing valid data. Real valid-data readiness remains separate from physical CELL coverage. Do not reject an existing legal physical payload solely for exceeding candidate valid extent. Preserve existing aggregate explicit-range, specialized zero-payload, packed/nonpacked and weight callers, atomic publication and existing source-stripe/destination-offset separation. Add actual singleton/multi-PE TMOV/TLOAD valid-tail witnesses and retain permuted offsets plus genuine missing-valid-data negative. Include the unchanged PTO-AVS-BLOCK-SHARED-EXPLICIT-RANGE-001 and exact affected existing regression/weight IDs selected without full-matrix discovery. Never alter the existing failing fixture to make the regression disappear.

   F17: finalize the existing authored source-backed TLOAD instruction projection and truthful EN/zh singleton-whole-valid-parent versus multi-PE stripe/independent destination-range summaries, current owner hash and meaningful existing site-contract assertions. Preserve first-fault/partial-data behavior and separate weight-specific rules. This is not an EN/zh reader-guide rebind. Previously accepted guide content remains protected; inspect actual binding/attestation rules after source changes and use the existing official review/rebind workflow only if required, retaining independent reviewer and distinct main verifier. New helper guide remains pending unless actually reviewed.

   Finalize all affected sources/ADR owner inventory/direct projections/census/traceability before clean signed candidate and exact prepare-pr packet. Re-run affected focused evidence and repository/site gates under their actual identity requirements; old fb4 77/77 and diagnostics are historical inputs, not automatic approval of a new HEAD. Confirm idempotence and no tracked generated build/cache. Main maps acceptance and final scope; same reviewer verifies F1–F17 closure, all changed and adjacent owners, truthful projections, compatibility and final passing gates, with one complete findings batch and fresh schema-valid exact-input receipt. Publication remains pending that complete acceptance. V1 only, no merge/release/downstream writes, no weakening validators or expanding architecture.

   Additional r2.13 owner/evidence closure: the existing PTO-B-ASSEMBLE-SHARED-STANDALONE-001 and PTO-TMOV-CONTRACT-001 / PTO-TILE-TMOV dependencies are now included in the same ADR inventory. Row-tail representatives cover encoded singleton TLOAD/TMOV physical4x16 with valid1x8 and full-mask TLOAD partially valid rows, including real selected participants with no valid source coordinates. These preserve the existing valid-region and whole-parent rules; physical coverage alone does not imply valid data or enlarge the logical valid rectangle. The current-source 16-point correction slice passes. Final signed input, exact required gates, independent complete review and publication acceptance are still pending. No new public architecture decision or release/downstream completion is claimed.

   25. Complete round4 correction batch (r2.14; existing-owner repair)

   Main read the same independent reviewer's complete review_complete=true report and exact schema changes_requested receipt for2ec158ff766e4222c0bb6fd53d5a702868753767, tree4bd95760b27486c76f189f4b9687ad7c8a0f7e42, diffSHA256d3e45263c6635191d067484278c04f9731f69ae6ee6230ead68d098e9bb237bb. Three P1 findings F18–F20 and four real production counterexamples are consolidated; report/receipt/map/input are preserved under build/issue360-corrections/round4-preserved, probes and six executions under build/issue360-review-round4/results.json. All prior contract sections remain binding. These are repairs to existing SourceExtent, descriptor bounds, final-data readiness and aggregate producer contracts, not a new public architecture choice. Same sole GPT-6 Luna XHigh executor and independent GPT-6.1 Sol Medium reviewer continue.

   • F18: actual TMOV's placed fragment logical extent must use declared valid coordinates within the actual copied source_elements/view. A larger legal backing source must not grow validRows beyond copied data or physical parent bounds. Preserve sparse physical positions and non-valid tails; do not substitute physical writer size or undefined-bit-derived validity, invent a source/writer equality precondition, or change faults. Real encoded fullmask larger-than-writer512B4x16/valid4x8 sources with four128B writers must publish a legal4row parent with the true copied valid payload. Protect singleton/sparse masks, permuted legal destinations and ordinary TLOAD whole-parent semantics. Owners: shared-generation.asl, shared-generation-pe-requests.asl and existing TMOV/descriptor clauses.
   • F19: final coherent descriptor readiness must prove every required valid coordinate from actual defined payload. Width/height expansion cannot reuse non-valid-padding ready bits from an earlier smaller descriptor. Recompute affected older proofs or provide equivalent complete final verification; no fabricated padding, blanket source-shape equality, aggregate bypass, or shrinking the final required rectangle to hide missing data. Real fullmask width8→16 and legalmask1100 dual256B height1→3 counterexamples must remain incomplete/unpublished while required coordinates are undefined, with old parent, ordinal and exact request/release bookkeeping preserved. Cover arrival permutations and existing packed/unpacked/zero-payload boundaries. Owners: shared-generation.asl, shared-generation-state.asl, shared-generation-pe-requests.asl; existing atomic whole-parent readiness contract.
   • F20: preserve authoritative caller representation, including aggregate cooperative weight's destination-placed4row fragments forming16valid rows, ordinary TLOAD's complete-parent rectangle, actual TMOV and specialized fragment producers, and generic non-valid tails. A context flag cannot classify every aggregate producer as complete-parent. Retain existing explicit placement and true data readiness with no blanket aggregate bypass or new shape/fault policy. Keep the existing failing tests/asl/block/model/dispatch/weight-to-shared-execution/block-state-tload-weight-cooperative-001.asl byte-unchanged and include exact PTO-AVS-BLOCK-TLOAD-WEIGHT-COOPERATIVE-001, plus affected aggregate/TIMG/weight/packed/zero regressions. Inspect direct callers and amend the SAME ADR inventory with existing affected clauses/units only where missing; no new ADR.

   Preserve actual red evidence before correction. New tests must assert corrected externally observable behavior rather than assert defect states as passes. Extend acceptance from the old90union with all four counterexample purposes and missing unchanged cooperative weight ID; count actual unique IDs only after selection. Owner-to-evidence map must name complete caller protection, without claiming selected Windows/KOrder/Equivalence prove cooperative shape. Finalize source before generated mirrors/catalog/evidence/source bindings, idempotent projections/census and clean signed candidate. Run current prepare-pr-selected checks, focused exact matrix, make pr-check, make repo-check and relevant site checks under eligible actual inputs. Earlier2ec90/90, pr-check11/11 and site/projection/Python passes are rejected-candidate history, not approval of corrected source.

   Main refreshed main at2026-10-09T07:46:47Z: e182 unchanged; #341 OPEN at950011003757f3dd6da811ef82c1891980a7dce8, unconsumed. Same reviewer must complete F1–F20 closure, all changed/adjacent owners/callers, compatibility/projections/gates and exact receipt. Unchanged prior conclusions may be reused with explicit immutable input mapping. Publication remains pending complete acceptance. V1 only, no merge/release/downstream writes, no validator/toolchain weakening. Inherited security13 FAIL, lost original baseline chronology, signature trust limits, pending helper guide and unassigned release obligations remain explicit residuals.

   26. Complete round5 layout correction batch (r2.15; existing-owner repair)

   Main read the complete review_complete=true same-reviewer report, exact changes_requested receipt and eleven execution records for64272e447289593a4f12327c2059e39a30dd4ac0 (ASLtree12ed8b377259d57cfdff66e9141c6455b3a13d76). Findings F21/F22 are consolidated P1 regressions of existing layout/packing/copy/readiness contracts, compatibility breaking on this rejected input. The four prior RowMajor counterexamples are independently closed, but current94/gates passes do not close other legal layouts. Full report/receipt/map/input are preserved in build/issue360-corrections/round5-preserved; fresh probes/currentmodel/commands/hashes/11results in build/issue360-review-round5/results.json. All previous sections remain binding. Same sole GPT-6 Luna XHigh executor and independent GPT-6.1 Sol Medium reviewer continue; no new public architecture, layout, shape or fault decision/ADR.

   F21: compute placed valid row extent in the authoritative logical-coordinate domain of the actually copied view and existing destination byte/CELL placement, respecting physical parent bounds and fixed-column semantics. TileLogicalLinearIndex returns layout-dependent storage/packed indices, not a RowMajor row-end: TileStorageIndex/TileLinearIndex and TilePackedLinearIndex are the existing mapping owners. A stored-index maximum divided by logical columns cannot infer rows for ColumnMajor, ZN/NZ, or row-local packed slack. Do not convert layouts to RowMajor, compact data, widen physical bounds, infer validity from defined bits, or change offsets/column rules. For complete offset0 writers, preserve source valid rectangle and layout exactly. Use existing index owners for bounded/placed views; escalate a concrete representational conflict to main rather than inventing it.

   F22: readiness proves the final required logical coordinate set through owning layout/packing-to-storage mapping. TileLogicalElementDefined consumes the storage/packed index, so physical-element DIV/MOD columns is not a universal validity predicate. Every required coordinate must be actually defined; physical padding/odd-column packed row slack need no fabricated value. Keep physical CELL coverage separate and retain complete-parent recomputation on descriptor expansion, atomic publication and genuine-missing-data parking. The same correct helper applies to actual-PE and generic/contextless producers; no blanket exemption or padding-only patch. Existing no-growth ColumnMajor case proves this correction is independent of F21.

   Required minimum real encoded singleton offset0 full-writer positive witnesses (matching PE3 mask/source/destination allocation, real source production and stop):

   • ColumnMajor512B physical4x16 valid1x8, all8defined: final physical4x16/layout unchanged, valid1x8, publish/consume once, no fault/extra events, source/private/old-parent/ordinal/peer release protections.
   • ZN2048B physical16x16 valid1x8, all8defined: same preserved16x16/ZN/valid1x8 publication.
   • NZ2048B physical16x16 valid1x8, all8defined: same preserved16x16/NZ/valid1x8 publication.
   • E2M1X2 RowMajor512B physical64x15 valid2x15, all30defined, physical row-pair slack undefined: preserve packing/shape/valid2x15, publish without requiring slack.
   • Independent ColumnMajor512B physical4x16 valid4x8, all32declaredcoordinates actually copied/defined and no row expansion: readiness true for required data, no false demand on indexes32+ non-valid columns; real release once.

   Preserve genuine missing-data negatives and add minimal relevant legal layout/packed placed-view or offset boundary evidence only using existing SourceExtent/CELL/index owners. Current Shared CUBE_M16/M32/N8 rejection before candidate remains protected; CUBE_N16 is unassigned and must not be introduced. Keep ordinary TLOAD whole-parent, generic FALSE/PE explicit !whole_parent_tload/weight-TIMG TRUE representation and old byte-unchanged explicit-range/weight fixtures. Current B.ASSEMBLE range/TMOV/layout/definedness/packed owner references remain authoritative. Amend the SAME ADR inventory/direct dependencies only for actual affected existing owners/clauses, no new ADR or generated normative prose.

   Preserve red fingerprints, actual fixture/model/source identities and the original source-coordinate evidence before overwrite. New permanent points must assert corrected externally observable success/data/bounds/readiness, not defective state. Extend prior94unique selection with each new representative purpose; count actual unique IDs and preserve all previous points. Source stable and minimum representative feedback precede final projections/signing; final semantic commit precedes commit-scoped census and receipt-only commit. Preserve unchanged supplementary canonical bytes/shards/attests; generated ASL regions must be current. If actual supplemental claims change, genuine same-reviewer bounded acceptance plus main verifier precedes official rebind. New helper guide stays pending.

   Final clean signed input/owner map/fresh prepare-pr, focused exact matrix, projection/idempotence, selectedPython/site and normalpr/binary gates remain required. Redirect final site commands to current-input raw logs where practical; any actual streamed transcription must be labeled honestly. Historical64294/Python136/site14/build/projection/repo7/pr11 passes are rejected-candidate evidence, no automatic new-input pass. Same reviewer completes every F1–F22/changed+adjacent owner/caller/layout/packing/zero/compatibility/docs/gate requirement and gives one consolidated verdict with fresh exact schema receipt. Main then makes actual owner acceptance before authorized ready PR creation. V1 only; no merge/release/issue closure/downstream writes.

   Freshness at2026-10-09T09:48:38Z: main actually fetched origin/main=e182c9b70d54a3a264bac54af9b915a0e43bb519 unchanged; #341 OPEN at950011003757f3dd6da811ef82c1891980a7dce8, unconsumed. Inherited security13FAIL, original lost red chronology, signer trust limitation, unassigned release and pending helper/downstream obligations remain explicit.

   27. Complete round6 parent physical-geometry correction (r2.16; existing F21 owner repair)

   Main read the complete review_complete=true round6 report, exact changes_requested receipt and all14 source/log-hash-matched execution records for179bed3276585bcd0e9fbb83f6a423e30ddd72cf (ASLtree183682dfa900b951177d29fed5668f2c6c8b4b87). One consolidated P1 F23 remains: INIT replaces actual legal physical rows with the DerivedTileRows capacity upper bound before raw payload copy. This is unclosed F21 full-view preservation, not a new layout/shape/fault architecture. Same sole GPT-6 Luna XHigh executor and independent GPT-6.1 Sol Medium reviewer continue. Fresh fetch2026-10-09T11:51:56Z confirms basee182 unchanged and #341 OPEN/head950011003757f3dd6da811ef82c1891980a7dce8/unmerged. No rebase required. All prior frozen constraints/owners/12criteria/F1–F22 remain binding.

   Rejected179 snapshot: build/issue360-corrections/round6-preserved/{review-report.md,pr-review.json,prepare-pr.json,acceptance-map.md}; strict14probes/model/commands/hashes/results remain build/issue360-review-round6/results.json. Main hash-verifies all14 source/log pairs. Correct new positives fail on179 at intended semantics, strict defective predicates pass; old642 also fails the same legal ColumnMajor instance. This is not an inverse-only new regression or invalid source-capacity diagnostic029. Actual17999/99, projection/Python136+1skip/repo7/site14+cleanENzh rawlogs passed; pr-check was NOT STARTED after confirmed rejection. None certifies corrected final input.

   Authoritative owners: asl/block/model/operands/shared-generation.asl common INIT/parent construction and raw-copy submission; shared-generation-state.asl existing correct inverse/readiness helpers and callers; asl/tile/model/shape/rows-columns.asl TileShapeMatchesCapacity and DerivedTileRows; legality/descriptor-shape.asl and allocation.asl; existing TMOV source/binding capacity preflight in block dispatch/shared-tlsu.asl; TileLinearIndex/TilePackedLinearIndex and SharedTileUpdateCompatible. Odd FP32/FP16/BF16/packed row-local profiles legally accept rows<=DerivedRows. DerivedRows is an upper bound for these admitted shapes, not mandatory actual physical geometry. Existing ADR-SCALAR-0007 and18-clause inventory suffice unless actual changed authoritative scope requires the ordinary inventory update; no new ADR decision is allocated.

   F23 required before→after: same-capacity complete copied view may no longer change its physical rows/column stride/shape at INIT. Identify the existing full-copy case from source capacity matching parent capacity, offset0, complete parent coverage and complete actual payload extent (including copied element span, independently of logical validity). Preserve source layout/physical geometry and exact valid rectangle with real defined data for this case. Let the common owner select that bounded representation; an explicit internal full-view fact is allowed if needed, not a software feature. Do not unconditionally parent.rows=source.rows for every candidate; partial views, different-capacity generic fragments, whole-parent TLOAD and specialized weight/TIMG retain their existing parent construction/geometry and bounds. Do not use valid extent/defined bits as a full-copy detector, add new faults/preconditions, relax same-capacity TMOV admission, convert layouts, compact payload or fabricate padding/data. Earlier F21/F22 coordinate inverse and genuine-required-data proof remain in place. Escalate a concrete representation conflict to main before inference.

   Minimum real encoded closure, add permanent positive regressions with exact <=64char unique IDs:

   1. PE3-only mask0001, FP32 ColumnMajor source/parent/fullwriter128B, physical8x3/valid1x3, offset0, all3 coordinates actually written. Existing schema must accept it. Fault-free STOP publishes once and releases/retires once, preserving128B/physical8x3/valid1x3/ColumnMajor and the three real values at original source indexes0/8/16. No new memory effects/private source corruption, false9x2 growth or wait.
   2. PE3-only mask0001, E2M1X2 RowMajor source/parent/fullwriter512B, physical8x15/valid2x15, offset0, all30 real values defined, odd row-pair slack undefined. Fault-free STOP publishes/releases/retires once preservingphysical8x15/valid2x15/512B/type/layout and real values/slack, not capacity-max64rows.
   3. Small protected slice before widening: byte-unchanged generic explicit-range smaller128B source→256B parent retains complete parent capacity/geometry; ordinary singleton/zero-stripe/permuted source-vs-destination TLOAD; old actual copied-view/nonzero offset, required-data missing row/column parks and READINESS020; existing cooperative weight/TIMG actual construction. Five024–028 layout positives remain required. Preserve source-capacity equality and illegal SharedCube boundaries. Representative feedback must close before broad finalization; no dirtyfull99 sweep merely to discover schema mistakes.

   Then final source edit→official projections/idempotence→signed semantic commit→committed-ASL census/receipt-only clean final→fresh prepare-pr/map/all12criteria/F1–F23. Union preservesall99 and addsatleasttwo permanent geometric positives; report actualunique count (expected101), not assumed count. Final exact runner/gates/projections/selectedPython/site/binary/pr and complete same-reviewer closure remain required; final gate logs redirected and exact commands/HEAD/exits retained. Guide supplements/shards/attestations remain unchanged if canonical supplementary bytes are unchanged; generated ASL-block changes alone do not trigger attestation rebind. Actual supplement/binding changes require real same-reviewer bounded approval plus distinct main verification then official tools, never fabricated acceptance. Helper guide remains honestly pending. V1 only; no merge/release/issue closure/downstream changes. Main owns authorized ready PR publication only after corrected exact input/gates/independent approval.

   r2.16 explicit producer representation clarification: working-source minimum2 and protected27 passed before the final parameter wiring. Main permits an internal preserve_source_geometry input distinct from placed_fragment: only the actually eligible encoded TMOV producer supplies TRUE from its authoritative operation classification; generic wrapper, ordinary whole-parent TLOAD and specialized weight/TIMG supply FALSE. The common owner still requires sourcecap==parentcap/offset0/fullparentcoverage/fullactualpayloadandcopiedelementspan; the boolean is not an unconditional geometry override. Do not infer TMOV from arbitrary live state, specialized_inputs_valid, or non-TLOAD unless complete caller ownership proves that equivalence. Completion-first, source legality and nonparticipant ordering remain unchanged. This is a permitted bounded representation of the existing contract, no new software feature or ADR. Fresh post-wiring sourcefeedback/finalization/exact focused/gates/review remain required.

   r2.17 — complete round7 mechanical evidence closure

   The complete independent review of clean037 found no remaining semantic defect after F1–F23 closure; its sole consolidated F24/P2 finding is stale generated architecture-readiness and release-gate-readiness JSON. The same executor will regenerate those artifacts through official dependency-ordered generators and finalize a fresh signed input, followed by the same independent reviewer and actual eligible gates. This is projection maintenance under existing owners, with no new architecture decision/ADR, release validation or downstream change. The previous current-input aggregate projection/pr/repo/Python/site successes do not override these direct freshness failures. Final implementation acceptance and publication are now complete for the reviewed input recorded below.

   Implementation completion — ready PR #373

   PR #373: Add ordinary SYNC.CORE with independent PE contexts is open as a normal ready PR (not draft). Published head: 47666ac37eb082faddff5073f0736b87c5a7e021, baseline/merge-base e182c9b70d54a3a264bac54af9b915a0e43bb519; head tree f508805e4521390fe9e0a34f7e96aa39961d5223, diff SHA-256 0ee97886cea0101d1181dae33e73d89a7f8b4740329a016600d5c34a43776f06.

   The same GPT-6 Luna XHigh executor completed APPLY, and the independent GPT-6.1 Sol Medium reviewer completed all 12 acceptance criteria and F1–F24, with verdict approved, no remaining findings, compatibility compatible. Main owner acceptance and prepare-pr --review passed for this exact input. ADR-SCALAR-0007 is committed; current meaning belongs to its named ASL/NDF owners.

   Local evidence: 101/101 selected ASL points passed on037; the final projection-only commit retains identical ASL/fixture/model/runner inputs, so this is explicitly source-bound reused evidence rather than a claimed execution on476. Exact476 Python136 tests/1skip, repo-check7, pr-check11, pinnedpnpm site typecheck/unit14/clean EN+zh build all passed. Official direct readiness checks, affected downstream projections and idempotence passed. Two final readiness JSON updates do not alter architecture semantics.

   Publication snapshot: hosted head equals the reviewed input; both PR source-contract and tooling-tests jobs are running. Hosted results and resolved review conversations remain landing requirements. Local agent receipt is not authenticated GitHub approval. Commit signatures are cryptographically valid, but local principal trust is unavailable and GitHub reports unknown_key for the published head. Inherited security13 advisory FAIL, lost original baseline per-ID fingerprint, pending helper reader guide, unassigned release/V2 validation and downstream adoption remain recorded limitations. #341 is still open/unconsumed and requires owner reconciliation whichever change lands later. No merge, release, tag, downstream write or issue closure was performed.
2. zhoubot

   added a commit that references this issue on Oct 10, 2026

   Add ordinary SYNC.CORE with independent PE contexts (#373)

   ...

   5641479
3. zhoubot commented on Oct 17, 2026

   

   zhoubot

   on Oct 17, 2026

   Contributor

   More actions

   Resolved by #373, merged on 2026-10-10 as 5641479c212abb86d46ac9cf7ececf82ac9e1a73.

   This implements the accepted r2 ordinary, operand-free SYNC.CORE contract: exact SYS envelope, independent PE contexts, four-PE rendezvous, earlier GM completion/ordering, interrupt/recovery handling and invocation lifecycle, with executable evidence and independent review. The accepted r2 decision supersedes the initial FENCE.D.CORE4/profile proposal.

   Closing the normative specification issue as completed. Downstream adoption and release qualification remain separate; the concrete platform launch ABI is still tracked by #386.
4. zhoubot

   closed this as completedon Oct 17, 2026