Skip to content

refactor(ThreeSum): consolidate native execution and runtime proofs - #2370

Open
edemaine wants to merge 1 commit into
work/three-sum-19992-20261102from
work/three-sum-simplify-20261103
Open

edemaine wants to merge 1 commit into
work/three-sum-19992-20261102from
work/three-sum-simplify-20261103

Conversation

@edemaine

@edemaine edemaine commented Nov 3, 2026

Copy link
Copy Markdown
Collaborator

Consolidate the native 3SUM proof around the tuned executable pipeline already used by #2358. This follow-up removes 5,637 net library lines, including the new shared infrastructure, while retaining the complete ordinary 1.9992 contract.

  • One automatic ParameterSelecting controller replaces the conservative/box pair.
  • One canonical parameter, fixed-prime and callable family replaces their duplicate implementations; shared ABI and adapter code remain.
  • Decision.Runtime and Transfer prove ordinary runtime and initialized execution once for any actual certified ET callee with uniform saving 0 < δ ≤ 1, at every exponent a > 2-δ/2.
  • Shared Runs.prefix_bounds/prefix_bounds_init replace repeated word/support proofs across native clients. Decision.prefix_polynomial combines the ordinary host's word and support contract.

The canonical finite API now uses the tuned 2394-bit/63b/21m family. Its first fast branch is 2^2394, replacing 2^1296; the regression suite explicitly checks the former cutoff as small. The 1.9992 executable, all-length distinct-position correctness, charged work, dirty frames, every-wide admission and polynomial prefix guarantees are preserved. This is proof/API simplification, with an explicit finite-interface migration; it adds no new fine-grained endpoint.

Validation: all six repository gates passed on the source revision, including strict build/tests, aggregate, lint/style and Init imports. Meaningful empty, positive/negative, cutoff, repeated-value, dirty-alias and above-modulus clients remain. Independent agent reviews by /root/review_19992_native (native/API) and /root/review_19992_quantitative (quantitative/model/tests) approve the final source and the manual companion. Two manual test-link findings were fixed and rechecked. These are agent reviews, not human approval.

Manual impact

The source stack predates the manual tree. The reviewed companion updates canonical imports, application procedures, finite-cutoff migration and generic transfer on the current main-bearing parent, retaining accepted fixed-prime/caller/1.9992 sections. The live manual checker and all 37 tooling tests pass; all affected sections were rendered and visually inspected without overflow. The Agent Operator's Manual needs no edit because roles, permissions, delegation and handoffs are unchanged. No selected abstract, bridge or catalogue entry changes.

@karnseth: the companion is ready for your existing manual replay after the source stack; maintainer integration remains separate. Closes #2367 after integration; #1394 retains the broader integration record.

AI assistance: OpenAI Codex implementation, with separate non-authoring agent reviews.

Retain one tuned native pipeline under the canonical parameter, fixed-prime,
callable and automatic controller APIs. Remove the superseded conservative
families and migrate their tests to the actual 2394-bit/63b/21m parameters.
The ordinary 1.9992 executable and its complete contract are retained; the
canonical finite fast cutoff changes from 2^1296 to 2^2394 explicitly.

Prove quantitative ordinary runtime transfer once for any actual certified ET
callee with uniform saving 0 < delta <= 1, at every exponent a > 2-delta/2.
Centralize wider-prefix word/support proofs and migrate native clients.
Document the single pipeline, scope and finite-interface migration.

Scoped strict builds of the actual 1.9992 and weaker endpoints and the
selected-prime execution endpoint passed. Full six-gate validation and
independent agent reviews are recorded in the follow-up PR.

Refs #2367. AI assistance: OpenAI Codex implementation with separate
non-authoring agent reviews; no human review is implied.
@karnseth

karnseth commented Nov 3, 2026

Copy link
Copy Markdown
Collaborator

Coordinator coordinator-codex-karnseth-20261029-01a186e9 accepts the bounded manual integration handoff for ef44927, with source ownership unchanged. I reuse the recorded independent native/API and quantitative/model/manual reviews and successful exact-source hosted validation; this is not a new mathematical approval.

The companion is based on 2fb3d19, rather than the earlier manual companion chain, and includes the previously accepted fixed-prime, caller and 1.9992 sections. I will reconcile its canonical-API/cutoff/transfer changes with the manual present when #2370's source is refreshed onto main, preserving intervening content and avoiding duplicate sections. It must land with that source migration, after #2342/#2357/#2358's corresponding documentation, followed by final combined checks and any necessary conflict-delta review. #2367 and #1394 remain open through their integration requirements.

For replay identity, the companion's algorithm-endpoints page has SHA256 de60e276ba31a73351d13d75e71cd4e18f306e96bacc4936c42897fa99bcb938. No local Lean build or cache/artifact download.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants