[AdaWorldAPI]

D-DRF-1: immutable DR64 focus → leased foreign cache → amortized dirty-row writeback

Status: DRAFT REFERENCE ONLY. This PR makes the operator's clarified architecture citable before any coding-agent prompt or EXO-50 handoff. It adds two research/preregistration Markdown files, and no Rust, ABI, wire, runtime, storage or CI code.

The gap this closes

• One DR64 focus is the normal case: an 8 KiB 64×64 u16 map into the existing 64K × 512 B SoA universe; maximum 2 MiB distinct referenced rows. DR256 global: 128 KiB pointer map, 32 MiB maximum distinct referenced rows.
• The 32×128-bit = 64×64-bit 512 B Boolean tile is not the u16 map. A separate 64³ Boolean relation cube is 32 KiB (40 KiB with the 8 KiB pointer map if both materialized). The third relationship dimension does not multiply canonical SoA rows.
• Sealed DR references are immutable and versioned. Cache eviction/reload must not alter provenance. Source SoA snapshots and physical row-index seals must also be pinned/witnessed.
• Loading a focus starts its residency lease. Lookahead both prefetches SoA facet cache lines and prepares the next Focus B. A's old readers finish safely; its disposable cache is released when the final dependent reference is gone.
• Legal pre-fold foreign gathering is optional and explicitly costed: 64 thoughts spanning Disease, Phenomenology and LOINC may reuse compact materialized facets instead of repeatedly chasing random references. No blanket no-copy doctrine; no silent repeal of existing ADMISSION B/H14c/H14d constraints.
• Accumulate a 512-byte dirty-local plane, translate once through the original DR version, coalesce by canonical row+owner using typed semantics, form final images against sealed Vn, and hand detached owned payloads to the existing async CycleWriter/WAL. Distinct facet updates must not be lost under naive last-payload-wins.
• Optional ONNX/Palette256 relationship plasticity is a separate candidate overlay. ONNX is not a native graph, certified proof, automatic learning rule or storage service; S3-like object storage is at most an optional versioned artifact-hydration tier, not a random-access page store.

Files

1. .claude/plans/d-dr64-focus-lease-foreign-cache-v1.md: ownership, geometry, lease states, prefetch, materialization, dirty-writeback and optional learning.
2. .claude/plans/d-dr64-focus-cache-falsifiers-v1.md: preregistered negative tests, A/B/C cost matrix and exact/approximate boundaries.

Explicit relationship to existing work

• D-RMAP-1A controls separate RelationshipMap and thought-palette ownership, generation witnesses and the pending pre-fold gather admission question. This PR does not add a map owner or redefine CE64.
• D-NFC / D-NFR / lab(mpf): D-MPF-1 — the Stage 1 math polyfill already exists; one expression, oracle, refusals, cost split #1752–D-ONNX-NATIVE-0: lance-graph-onnx, ONNX names over existing native functions #1754 control native-first method binding and ONNX as a downstream leaf. This PR does not add new kernels or an ONNX interpreter.
• docs(dr3) S3a–S3c: Cartesian reference, holographic focus, DR256³ validation plan #1747 / docs(onnx) S3d: Tensor × DR × Hybrid crossover + compiled Palette quality arm #1748 already document DR64³ and physical crossover. This PR specializes the missing focus memory lifecycle and dirty writeback, not another cubic ABI or benchmark winner.
• lab(D-NFR-6): native operations over the SoaEnvelope borrow — no Table, no frame guard #1755 separately probes real NodeRowPacket borrowing with ndarray. Do not recreate its withdrawn synthetic FrameStore / runtime guard.
• BatchWriter / cycle_writer / persist_sink / LanceCycleWriter retain sole-writer semantics. try_submit moves owned work into a background queue; the actual frozen payload/Arrow writer still has copies. Successful submit != durable commit.

Acceptance and exclusions

• This is a reference contract and falsifier preregistration; no benchmark or test run is claimed.
• Follow-up probe PRs should be staged: (P1) two-focus leased handoff, (P2) 64-thought direct/prefetch/gather/writeback cost and parity, (P3) typed relationship cube plus dirty reduction, (P4 optional) ONNX serialized candidate plasticity.
• No production interface, no new CE64 fields, no storage lease policy, no per-fold new guard, no adoption claim for ONNX learning, and no EXO-50 post or agent prompt in this PR.

Review request: confirm operator semantics and admission/ownership boundaries, especially dataset+branch versioning, compaction seals, scoped pre-fold gathers and alias-safe dirty-row reductions. Those should be settled from this PR before any code is commissioned.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 19eeeeb2-f1c7-4e13-b3d8-68088fbcdb15

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-24T05:01:08.729677Z	d16be83	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.