[AdaWorldAPI]

Lab only. No production crate changes. Please don't merge automatically. S-INGRESS and D4 are still open, and nothing here decides them.

Start with .claude/probes/2026-11-03-d-global-where-p1/REPORT.md. Two documents back it up:

• ENGINEERING-PARITY.md: the literature-first matrix.
• LANCE-TRACE.md: the caller path, traced against Probes: Lance 13 source pruning (D-TELEPORT-2) and 64 resident populations (D-PHOTO-0) #1489, D-HOLO H14c: thought-first positional reads, Hamming presence and Lance masked I/O #1573, D-HOLO H14d: logic → address mask → masked read → terminal (C only), measured on v4 #1575, D-SRAM-2 spike: masked Lance 13 reads and the addressed semijoin #1633, docs(D-DQM): DR64 ↔ DuckDB ↔ Lance 13 capability crosswalk (follow-up to #1787) #1788 and probe(D-DQM P1): chunk-touched predictor vs H14c/H14d Lance bytes — FAIL #1789 with source line references.

What goes to main (2,518 lines)

• Docs: REPORT.md, ENGINEERING-PARITY.md, LANCE-TRACE.md, SPEC.md, and the board entry D-GLOBAL-WHERE-P2.
• Lance arm: crates/crepe-plan-probe/examples/where_lance_arm.rs. It compares a full scan, an allow/block prefilter and the D-SRAM-2 spike: masked Lance 13 reads and the addressed semijoin #1633 attested range read, each with ClassView-driven column projection.
• Lab crate edit (crepe-plan-probe/src/addressed.rs):
  • ranges and range_mask are now pub.
  • New range_tree. The example builds exactly one mask tree per read with it, and measures the tree's bytes outside the timer.
  • New check_fresh; current_lane now calls it, so there is one freshness rule.
  • The example reuses these helpers instead of copying them. No new read API.

What stays off main

The research branch research/d-global-where-p2 (commit 1422b2b92) holds:

• the in-memory probe (probe/: falsify_p1, where_k, project, compose, corrections);
• the DuckDB 2.0.0 harness (duckdb/);
• all raw measurements (results/).

Each doc on main says that its probe/, duckdb/ and results/ paths refer to that branch.

Results

verdict	finding
Semantic parity: reached	Every answer equals the oracle and DuckDB at 1M and 10M. Every Lance arm equals the oracle and the full-projection oracle at 1M and 10M.
Performance parity: split	Count wins, from 14.2× down to 1.96×. EXISTS loses at 10M, 0.43× and 0.16×, because Terminal::Any has no per-tile early exit.
Engineering parity: reached on representation, not on strategy	Missing wiring: EXISTS early exit, and routing once per frame. Missing capability: choice of evaluation direction, and an ordered u8-in-u32 compare word.

Lance findings:

• Bytes follow page locality, not selectivity.
  • A clustered 0.1% read is 112× faster than a full scan at 10M.
  • A scattered selection reads 33–100% of the bytes, and at 10–50% it is 1.4–11× slower than a full scan.
• ReadPolicy::DEFAULT decides on covered rows alone. It sends a single contiguous 10% range to a full scan, which is 3.8× slower at 1M and 4.6× at 10M. Recorded only; the policy is not changed here.
• ClassView → column resolution goes through predicate_iri. It refuses AbsentField, Unauthorized, UnknownProperty, AmbiguousProperty and StaleSchema, each paired with a success control. An ordinal-mapping trap and a label-mapping trap prove a shortcut would be caught.
• BLOCKED_D4: arm C takes its ids as given. It composes no DR ordinal with _rowid and builds no third coordinate map.
• RowAddrMask ingress is a measured copy, never zero-copy.
  • It costs 0.4–5.4 µs for one range and 123,162 µs for 3.26M ranges.
  • These are the figures after the Codex fix. The first run timed two tree builds per C read and doubled them.
• Lance 13 truncates max_rows_per_file with as u32 (write.rs:1077), so 1 << 40 becomes 0. The D-SRAM-2 spike: masked Lance 13 reads and the addressed semijoin #1633 NotContiguous guard caught this at 10M.

Each verdict item in REPORT.md §8 is written as producer → missing edge → consumer → test → measured cost → owner.

Relation to #1949 (D-DRAM-0, documentation only)

This PR measures what #1949 plans:

• Family L (locator reuse): C1 shows that routing once pays from Q ≥ 2. The stale guard refuses a binding after replace_tile, and turning the guard off makes the stale binding pass.
• Family A (Abakus thresholds): C2 shows there is no ordered u8-in-u32 compare. Composing the existing ternary words costs one pass per 2-bit prefix, up to 5× slower than a hand loop. This bears on docs(D-DRAM-0): DR locator memoization and adaptive Abakus mask lens (architecture gates) #1949's unverified kernel signatures.
• Family C (DuckDB engineering parity): see ENGINEERING-PARITY.md.

Testing

• On the research branch: falsify_p1 37/37. where_k, project, compose and corrections match the oracle at 1M and 10M. The DuckDB parity scripts pass at 1M and 10M, threads 1 and 4.
• where_lance_arm passes every oracle and refusal check at 1M and 10M, re-run after the one-tree fix.
• crepe-plan-probe: cargo fmt --check, cargo clippy --all-targets -- -D warnings and cargo test pass locally (59 tests).
• Board checks pass locally: the entries index (569 rows), the supersession index, and citation decay (no new decay since base).
• Warm timings of 1000 µs or more are printed as FAILED-GATE. Storage, attest and prepare times are labelled COLD/IO. Every Lance byte count came from the page cache.
• Not run: the full workspace test suite. Only the lab crate and the docs change on main.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XAqKMNZqKXSaeb9g2ypywF

[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: e95e7b43-d06e-486c-89e0-f6b0fec6a8f9

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

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

[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 2026-11-03T22:30:28.948277Z	9bb2c91	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.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9bb2c91b1e

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".