[AdaWorldAPI]

Lab follow-up to #1790; it does not extend #1790. The lab lives in the workspace-excluded drf-focus-probe crate (examples/dr256_on_demand.rs). It was pre-registered in commit e01d38879 before any run, and results are appended to .claude/board/entries/2026-10-26-dr256-on-demand-l1.md. No production code changed.

Depends on AdaWorldAPI/ndarray#377 (mask_scatter_or_u16). The lab crate patches ndarray to the sibling checkout.

What it measures

Each thought brings its own 8 KiB global plane X and asks one of two questions over the #1790 fixtures: any(G ∧ X) or |G ∧ X|. Five arms answer the same question, each with its preparation charged:

arm	what it does per thought
A	local only, no global mask
B	rebuild G with the u16 scatter
C	rebuild G with the u32 scatter (widening charged once)
D	reuse G while its witness holds
E	count without a shared G

N ∈ {1, 2, 4, 8, 64}, run twice.

Findings

• Correctness: all masks and answers equal a set oracle, with whole-mask comparison. Hot-path allocations after setup: 0.
• Reuse: it pays from N = 2 on every fixture. At N = 64 the cost per thought falls to 0.3–0.5 µs (clustered, scattered) and 2.3–4.5 µs (chaotic, full).
• Rebuild per thought: at chaotic/full density it reaches milliseconds at N = 64, which is a failed gate under the workspace rule.
• Without a stable window (the selection changes every thought), reuse gives nothing.
• E is slowest for count.
• u16 vs u32: u16 removes the 1 MiB widening (9 vs 46 µs at N = 1).
• A answers any in nanoseconds only because X is 50 % dense and the walk exits on the first hit. With an empty selection A costs 2.4–4.9 µs per thought, against D's 0.3 µs.
• L1 interference with a resident DR64 fold (512 B mask + 16 KiB column):
  • Keeping one 8 KiB G live does not slow the fold: within 1.5 % in both runs.
  • Rebuilding per thought does: by 11.8 % and 13 % on clustered. No hardware counters were available, so this is timing only.

Decisions

• mask-risc exposure: designed, not implemented. The design adds LaneRef::U16, read by the existing ScatterOrU32 / ScatterCountU32 with the width resolved per tile; no new terminal. It waits for ndarray#377.
• Reusing a scattered G across thoughts is not lawful today (ir.rs:329-337). The entry records a narrow "witnessed derived population" contract proposal for your decision: read-only folds over the same universe, a full witness (source, scope, universe, every map's id and seal, every register's selection generation), bounded to one compatible thought window, never persisted.

---

Generated by Claude Code

[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: 0629c98b-62bb-4b54-a61a-2ecc4a4065d6

• 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	🔄 Running since 2026-10-26T17:48:33.647243Z	6323b35	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.