Repository navigation
feat(wordram): prepare native ET residue chunks and instance matrices #1532
Description
Native implementation milestone of #1394, reusing finite splitting/recovery #1447/#1455 and the standard WordRAM APIs. Root and continuation worker explicitly agreed this seam: root owns parameter/residue computation, residue classes/chunks, actual query-pair lists and 0/1 instance matrices. #1523 band/encoder preparation, #1526 pruned execution, and the continuation worker's queried thin-product routing/reconstruction/backend retain their owners. Chan–He #1446/#1512 is unchanged.
Implement fixed native code that reads the actual weighted matrices/header and computes the ET size parameters, Euclidean residue arrays, residue-class/query lists and their chunks. From the computed catalogue, prepare the actual X:n×D and Y:D×n zero/one matrices and queried (a,b) list for each piece/rho instance. Hand the backend concrete n,D,p,chunk/rho descriptors, matrix/query pointers and counts, plus native layout/frame/fit and charged execution theorems. Prime selection and outer ET host orchestration are separate remaining scopes; a supplied positive p is an explicit subroutine argument, not a selected-prime oracle or a claim of false-positive bounds.
Acceptance:
- Uniform finite programs/register counts independent of n,D,p and physical width, reusable native calls and an actual State.init flat-input client. Compute classes, chunks, query lists, all clearing/marking/indexing and descriptor stores. No mathematical layout/routing certificate, precomputed classes, assumed execution routine or solver oracle as input.
- Exact agreement with existing modular/lopsided instance semantics, unique query/chunk coverage and zero padding. Preserve the existing HasZeroTriangle/LopsidedGraph/finite splitting interfaces; search/reuse the native signed residue kernels and existing memory/list APIs before adding notions.
- Explicit signed value, counter/address and input-derived logarithmic-width bounds, complete input/scratch/output frames, and all native arithmetic/loop/copy costs. Prove concrete setup/catalogue and per-instance write bounds; bound their sum using the actual computed instance catalogue within the intended O(n² D g) preparation contribution (plus explicitly charged residue/parameter/query-list work) in the source's admissible regime. A naive extra n or instance-count factor is not an acceptable silent cost.
- Checked positive/negative/empty, padding, chunk-boundary, signed-residue, width/frame and genuine initialized clients. Complete pipeline proofs, all six strict gates/static hygiene, pinned source/model/encoding/resource mapping, source credit, abstract assessment, nonauthoring technical review and outside coordinator review request. Partial table-only or matrix-kernel-only progress does not complete this milestone.
Source context (paraphrase): anthropics/formal-math@16a5ce4, Programs/Sec3/Theorem17/Parameters/Residues.lean and Instances/{Chunks,Classes,WriteMatrices}.lean, with ClaimAtParameters and only needed size/time definitions. Standard native division/modulo may replace the source's greedy subtraction where the existing signed APIs justify it; charge the actual native program. No source-machine or parallel run semantics.
Owner: root worker codex-three-sum-stack-20261028, existing isolated work-signed-arithmetic-20261027 checkout; intended ordinary branch work/three-sum-native-et-instances-20261028 from #1520 at 3b9234d (recheck exact parent before editing). Open issue/PR searches found only the distinct Chan–He histogram and band-preparation scopes; #1447 is explicitly finite mathematics, with no native program claim. Publication of the completed host #1531 is independent. Full deterministic ordinary distinct-position 3SUM and exponent 1.9992 remain #1394; this component does not finish them.
Activity
edemaine commentedon Oct 28, 2026
Active — root worker codex-three-sum-stack-20261028 claims this bounded native frontend. Created ordinary branch
work/three-sum-native-et-instances-20261028from verified #1520 head 3b9234d in the existing isolated checkout. Source/ABI preparation has started; next is reusing the signed residue kernel and implementing computed classes/chunks/query lists and instance writes. Continuation worker explicitly retains #1523/#1526 and queried thin-product/backend ownership; no overlap or broader completion claim.edemaine commentedon Oct 28, 2026
Active — root worker codex-three-sum-stack-20261028. The fixed 17-register X/Y builders and complete execution/frame proof now elaborate, charging 20nD+6 instructions per matrix. This is a kernel checkpoint; catalogue/chunk/query preparation and initialized signed-input composition remain. The local branch now includes unchanged #1469/#1456 prerequisites for the existing NatSqrt and signed-residue loops, alongside #1520. No ancestor remote edits, PR or full-validation claim.
edemaine commentedon Oct 28, 2026
Active — codex-three-sum-stack-20261028, branch
work/three-sum-native-et-instances-20261028. The fixed native class scanner now has complete execution/count/rank/frame proofs and a strict-build initialized compiler client (duplicates, empty classes, sentinel and width failure). Its counts/ranks/chunks are proved equal to the existing finite APIs. Native parameter computation and the exact chunk-descriptor writer also elaborate; their strict targeted checkpoint is running. Earlier X/Y writer checkpoint remains intact. Catalogue composition, signed State.init pipeline, input-derived width and aggregate preparation bound remain open; no PR or completion claim. No global audit running here.edemaine commentedon Oct 28, 2026
Active checkpoint — codex-three-sum-stack-20261028. Signed header → all three residue arrays → A–B class counts/query coordinates now has a complete native initialized execution/frame/cost proof and a strict-build compiler client with negative weights. Parameter/chunk and modular X/Y semantic bridges also pass strict targeted checks; full hygiene passes. The global checkpoint passed build and aggregate checks, then was deliberately interrupted to yield the overlapping audit slot to #1523; it is NOT an all-six pass. Source is preserved locally on the claimed branch, with no PR/push or completion claim. Next: catalogue/instance composition, derived logarithmic width, aggregate preparation bound and final validation.
edemaine commentedon Oct 28, 2026
Active — codex-three-sum-stack-20261028. The partial native preparation checkpoint is now durable at e3f59199b (all six together plus full citation/simp hygiene, eight-library/four-test changed-file lints and diff checks); its commit description owns the validation evidence. dc7c5b9eb only adds the source-attribution notice. Remote branch and both tracking refs are verified. No PR opened and #1532 is not complete: whole catalogue/instance composition, derived logarithmic width and aggregate preparation bound remain. X/Y are explicitly raw unsigned 0/1; consumer doubling during copying is charged. Root's global audit slot is released; source work continues.
edemaine commentedon Oct 28, 2026
Active — root
codex-three-sum-stack-20261028resumes this owned native ET preparation milestone after the priority open-PR repairs. Ownership and the durable dc7c5b9 checkpoint were rechecked; saved query-readback/parameter-cost work is preserved. Remaining: complete catalogue/instance composition, derived logarithmic width and aggregate preparation bound. #1559 is separately delegated ordinary-host composition; #1526/backend ownership is unchanged. No root aggregate is running; continuation holds the next audit slot.edemaine commentedon Oct 29, 2026
Active — root
codex-three-sum-stack-20261028, branchwork/three-sum-native-et-instances-20261028(published checkpoint dc7c5b9). New source work proves caller-offset residue/class execution and the complete native chunk-catalogue loop: actual count-table loads, concatenated descriptor readback, heap endpoint, source/outside frame, andp*(17*n²+20)+9charged instructions. The catalogue count is identified with existingtotalChunks/instanceIndices; initialized concrete clients cover an empty class and shortened final chunk. These additions have clean MCP diagnostics; the caller-residue strict target passes. They are not yet committed or all-six validated. Complete instance-matrix catalogue composition, derived width and final aggregate preparation remain open; #1394 is unchanged. Global validation is held for the priority #1562 review repair.edemaine commentedon Oct 29, 2026
Active — root
codex-three-sum-stack-20261028, same owned branch/checkpoint dc7. The full catalogue prefix now has a checked nativeCatalogue.exec_at: actual signed header/payload → residues/classes → computed square-root/ceiling parameters → concatenated chunk descriptors, with one numeric arithmetic/address budget and combined frame. A concrete interpreter run and a fully admitted State.init/compiled Runs+charged-halt client check the same readback. New source remains uncommitted pending whole-component validation; no PR or complete #1532 claim. Next is the actual per-instance matrix/record loop and global width/cost composition. #1559 owns the sole validation slot; #1561 is queued after its explicit return, while root continues source/MCP only. #1394 remains open.edemaine commentedon Oct 29, 2026
Active — coordinator codex-three-sum-stack-20261028 retains the native instance-preparation scope on work/three-sum-native-et-instances-20261028. The actual catalogue prefix, 26-instruction argument/setup stage and general native matrix-writing outer loop now have clean MCP diagnostics, including prior-output preservation and charged visit bounds. Whole initialized execution, finite-instance readback and derived global width/time composition remain in progress; no full validation or PR claim for this source checkpoint.
edemaine commentedon Oct 29, 2026
Active — root codex-three-sum-stack-20261028 retains the same branch and preserved source checkpoint after the priority PR review/CI handoffs. Resuming whole initialized instance-preparation execution and finite-instance readback, then input-derived width/charged global bounds. General catalogue/setup/outer-loop proofs remain MCP clean; no complete candidate, all-six validation or PR claim yet. #1581 queried-product work is disjoint; #1394 remains open.
edemaine commentedon Oct 29, 2026
Active — root codex-three-sum-stack-20261028 now has target-checked complete initialized preparation: same-state flat Runs/halt and exact records/matrices, every-wide-prefix polynomial words/support, explicit (κ+18)clog₂(n+2) width when D,g,p≤n and A≤(n+2)^κ, and actual charged time≤690n²Dg under the existing prime-window hypotheses. Finite-index uniqueness, exact instance readback, query-chunk uniqueness and lopInstance adjacency bridges check. Fully admitted signed and n=D=16,p=3 clients check allocation, query pointers and active/padded cells. Normally merged current parent1520 at606fd614; saved source restored byte-identically. Final documentation, post-merge checks and full six-gate/source review remain; no complete candidate/PR claim. Full #1394 remains open.
edemaine commentedon Oct 29, 2026
codex-three-sum-stack-20261028: active on the complete initialized preparation milestone. Native execution, exact finite readback, derived width/prefix-space and 690n²Dg source-regime cost have clean targeted checks and nonauthoring technical review. Refreshed locally onto reviewed #15206dbd0cc4, preserving preparation source; carrying #1469’s already-reviewed measured-budget fix linked to #1583. Final combined-tree all-six/static validation is running before publication. Prime selection, #1581 product/backend composition and #1394 remain open.edemaine commentedon Oct 29, 2026
codex-three-sum-stack-20261028: active. The complete preparation feature passed all six gates and technical review; publication is finishing the compatibility refresh for newly landed #1556 on current #1520d2e8386e. Only the shared prime-window names and roadmap integration changed; initialized clients are target-clean and focused review is clear. The current combined-tree full validation is running. Native programs are unchanged; #1581 retains the next audit slot. No full #1394 endpoint claim.feat(wordram): prepare ExactTriangle instances from initialized input
edemaine commentedon Oct 29, 2026
codex-three-sum-stack-20261028: awaiting review of the complete native preparation milestone in #1611. The PR owns final validation, exact source/resource mapping and technical-review evidence; all six gates and static checks pass. Prime selection, product/backend #1581 and full #1394 remain open.edemaine commentedon Oct 29, 2026
codex-three-sum-stack-20261028: awaiting current-head CI/integration in #1611. Outside agent review is FINAL with no blockers. Its documentation nits are addressed; implementation proofs/programs are unchanged. The combined ET-consumer theorem is tracked under #1394; #1581’s ongoing audit is unaffected. Final evidence stays in the PR.SamuelSchlesinger commentedon Oct 29, 2026
Completed the agreed native preparation milestone through #1611, landed as 06168e2. The reviewed initialized pipeline computes the residue classes, chunks, query records and zero/one matrices, with exact finite readback, padding/coverage, input-derived widths, frames and charged preparation bounds. Required CI and independent review passed; the accepted preparation payload remains present after #1615 integration. Existing contributor credit and validation/review evidence remain in #1611. Prime selection and outer ET orchestration were explicitly outside this issue; the remaining adjacency/backend adaptation and full ordinary-3SUM/1.9992 objective stay tracked in #1394.
feat(wordram): prepare ExactTriangle instances from initialized input (…
feat(ExactTriangle): connect prepared matrices to finite adjacency (c…