D-HOLO H14c: thought-first positional reads, Hamming presence and Lance masked I/O

[AdaWorldAPI]

What this measures

H14 and H14b asked how cheaply each layout rebuilds a row. Thought-first C does not need rows for ordinary computation, so H14c asks a different question: once the thought, its operands and the Cartesian address are known, which accesses are still necessary?

Thought(t)[Table(a)[i] : Table(b)[i]]

The thought's program selects the lane tables (handles resolved once). The address i selects the operand in each table. A presence bit says whether a value is there, never where it is. Arm M from H14b is withdrawn as a layout candidate (#1569).

Pre-registered before any benchmark: 47b087e6 (.claude/plans/d-holo-h14c-thought-first-v1.md). Results were added after it; only the status line of the pre-registration changed.

What changed

• crates/lance-graph-mask-risc/examples/holo_thought_first_probe.rs:
  • a typed read path: Selection → Resolved handles → read(i) → Overlay / Base / Absent / Denied operands;
  • presence populations as mask-risc And/Xor/Or + Count programs, with no new arithmetic;
  • an explicit terminal row;
  • arms C (lazy and eager positional tables), B (connected delta with rank) and S (sparse, binary search);
  • the lane_footprint(&Program) instrument;
  • 11 oracle tests.
• crates/holo-lance-masked-io-probe (new, excluded from the workspace; its only dependencies are lance 13, arrow 58 and tokio):
  • Lance 13 with_row_addr_prefilter reads, combined vs independent vs unmasked, read statistics from scan_stats_callback;
  • a CoordMap (coordinate → _rowid at one version) that refuses a stale version;
  • 5 tests: equality with the reference, read order, identity across rewrites, the stale-map hazard, missing lanes.
• The workspace Cargo.toml gains one exclude entry. No dependency of any workspace crate changes.
• Plan, board entry with the raw CSV, STATUS_BOARD row, INTEGRATION_PLANS, AGENT_LOG (one Haiku census run, spot-checked).

Results (v4, tier: v4 (avx512f), rustc 1.99.0, lance 13.0.0)

	criterion	result
A	operands read directly, no row-first discovery	met
B	tables the thought touched but did not select do not add work	met: tables consulted == k for U = 2, 8, 31; at k = 2, U = 31 costs 0.90–1.01× U = 2 hot and 0.95–1.16× cold
C	populations without loading values	met: 0 value loads; 0.9–1.2 ns per 64-address word
D	no row materialization outside the terminal	met
E	right coordinate across rewrites	met for the checked path; every stale case is refused

• C vs B on query-shaped reads is mixed. Hot, B is faster up to 1 % dirty: B tests one row index for all k lanes, C tests k presence planes. Cold, C is faster from 10 % dirty at k ≤ 8. At k ≥ 16, B is faster in most cells.
• Memory. Positional tables are resident by touched page. Uniform 1 % costs ≈ 955 KiB per table for ≈ 10 KiB of values (≈ 93×, flagged at every T). Clustered 1 % costs 62 KiB. Seal is page-fault bound at low density.
• Lance 13. Both halves of the prediction hold:
  • uniform 1 %: the mask reads the same bytes as no mask (ratio 1.00), while rows scanned drop to 692 of 65,536;
  • clustered 1 %: 0.24× the bytes;
  • the materialization style changed bytes in 0 of 120 cells.
• Identity. A stale map gives silent omission, not wrong coordinates. Stable row ids keep _rowid across update and compaction (0 moved), but not across overwrite.
• Workload harvest. Over test programs only (not a workload): 169 Program::new, 61 literal; distinct lanes 0:26, 1:31, 2:4.

Disable runs (each turned its test red; files restored):

• drop the presence test;
• shift the address by one;
• substitute an unrelated lane;
• probe every touched table;
• build the row inside the point read;
• treat absent as zero;
• skip the aperture;
• remove the version check (both identity tests).

Not measured. This VM exposes no hardware counters, so cache and TLB behaviour is read from hot vs cold timing and page faults. Lance does not expose decoded bytes, pages accessed or allocations.

Open, returned to the operator

• How to bound positional-table residency without reintroducing traversal.
• Whether lane datasets should enable stable row ids.
• Whether data read by scattered masks should be laid out so presence is page-local.

No layout is decided.

Links:

• H14: D-HOLO H6c results; H14 pre-registered (full-row writes vs connected lane deltas) #1566
• H14b and the M withdrawal: D-HOLO H14b (withdrawn): positional lane mask M; measurements kept, no layout verdict #1569
• Lance Scanner: https://docs.rs/lance/latest/lance/dataset/scanner/struct.Scanner.html
• row-id lineage: https://github.com/lance-format/lance/blob/main/docs/src/format/table/row_id_lineage.md

🤖 Generated with Claude Code

https://claude.ai/code/session_011BMJnDBAfzcuQh5HeT39Rh

---

Generated by Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

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

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 57524079-48b1-4255-924c-10e193931c83

• 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.}

[coderabbitai]

Add Carrot credits or activate Agent usage billing to use Autofix

[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 2026-10-15T23:50:07.295375Z	39488c5	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: 39488c5755

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".