Skip to content

Complete selected BIT node for 64 <= n < 128 - #2697

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
jmbarrancoml:research/selected-node-caller
Oct 16, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
jmbarrancoml:research/selected-node-caller

Conversation

@jmbarrancoml

Copy link
Copy Markdown
Contributor

PR2569 supplied raw source packets and owner routes, but its consumer still lacked a complete selected node. This contribution constructs the node for 64 <= n < 128: original-offset remainder peel, fresh-row decoding, the full local/idle/bridge/exchange caller, an independent external gap and return to the original row order. All its paid children fit the completed below64 leaves.

The written refinement covers arbitrary dirty, correlated payloads under the pinned generated-program and source contracts. I charge the caller's actual 661-role staging, directories, per-role family scans, temporary outputs and completed leaves. Under the retained root coupling, I obtain a constructive C_q V(n+w) time bound. Numerical reserve admission, larger recursive returns and full multiplication transfer remain open; I leave the selected exponent unchanged.

I preserve PR2234's ordered 28278-request catalogue and distinguish native idle headers from common headers after owner routing. I also enforce two state invariants: scatter repeated-family roles into the latest family state, then finish one common low before gathering the next. The negative controls retain intermediate checkpoints because repeated faulty sweeps can cancel at the final endpoint.

Validation uses standard-library CPython 3.12.13:

  • Registered source, node and optimized-control suites with exact receipts and input hashes.
  • Full selected catalogue/ownership bindings and 7843392 literal events across 288 calls.
  • Complete GL2 node on 31104 independent Boolean columns, with 672 residuals and 5712 child entries/returns; complete gapped endpoint on 43740 bits.
  • Actual selected common/native coefficient junctions and generated width65 peel; sampled two-high-field execution at width2; 16 rejection controls under -O.

The complete Boolean executions use small models. I do not claim full GL64 payload execution or formal-kernel verification. The package contains proof, source pins, receipts, registered commands and structured next questions. I retain 25 predecessor files byte-for-byte.

I credit ywz65's conditional staging PR2475, primitive audit PR2566 and cubic directory-setup correction PR2590, plus the full source lineage in NOTICE.md. I use PR2569's corrected proof at 898ff36fe2cd05f733de6b26c88a4c4741b06237; its runtime bytes remain unchanged. Jose Manuel (jmbarrancoml), with substantial OpenAI Codex assistance.

Coordination and revision5 expansion decision: #2319 (comment)

I fast-forwarded to canonical main b5ea56e0a5ac4552166253b1fa4be0753bb38c30; priority remains 5 and the scientific pins match. The final overlap refresh covers 782 open PRs, 1320 Discussion/comment/reply records and 1690 local records, with complete pagination and no warnings. Supplemental reads through 17:40UTC confirmed unchanged main and PR2688, with no new node claim. New PR2688 supplies a complementary directory-only reserve component; it leaves total caller workspace open.

Reproduce:

python3.12 scripts/research_package.py --package research/selected-bit-node --check-only
python3.12 scripts/research_package.py --package research/selected-bit-node

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