board: catch up #1621–#1626, one row per duplicated blocker, D-RMAP-2 AMX re-run

[AdaWorldAPI]

Board only; no code change. One commit (a2286d2) off main bc8012da.

1. Arc rows for #1621–#1626

PR_ARC_INVENTORY and LATEST_STATE stopped at #1620. All six PRs carry code, measurements or plans, so each now has an arc entry, and LATEST_STATE gains a "main after #1626" section. #1621's own body asked for its arc entry after merge.

2. One issue row per duplicated blocker

Each of these blockers was recorded under several names. Each now has one ISSUES.md row that lists those names with their locations:

• ISS-ADMISSION-PREP-BINDING: may preparation translate global identities into focus slots before a fold? The same question appears as:
  • the D-RMAP-1 §6 BIND review;
  • D-MPS-8 "not admitted";
  • A1 in D-RST and in D-RPR;
  • D-RPR-1's wait;
  • GATHER-DECISION in ISS-GATHER-HOP-1616.
• ISS-RECURSION-DRIVER: no lawful recursion driver exists. Recorded as D-SRAM-4, D-MTP-3, D-RA-4's d′, the D-RPR loop, and the JC-SBI note. chain::ask from Consolidate into #1622: D-RPR-1 sub-cases, plan links for #1601–#1617, and the handoff lab probe #1624 is cited as one candidate shape, not as the driver.
• ISS-RETRACTION-PROVENANCE: per-derived-cell provenance is undecided. Recorded as D-TGA-3, D-RPR-2, D-RMAP-1A X9 and D-RST-5.
• ISS-DURABLEREF-THOUGHTCOMMITID: both types appear in crates/ only in doc comments. Owned by D-RST-1, coordinated with CPG-00.

The existing rows keep their own wording.

3. Smaller entries

• Correction: the 2026-10-19 plan-index line says D-RMAP-2 is "not run". A dated correction entry supersedes it; the file is append-only, so the old line stays.
• AGENT_LOG: the Haiku runs behind this and the previous synergy map. The board scout's claims all held. One count was wrong: the code scout reported 21 MaskOp::Gather matches, and a recount found 57 lines. No card is retired.

4. D-RMAP-2 §5a-AMX re-measured on a second AMX host

New entry 2026-10-20-d-rmap-2-amx-remeasured.md.

Host and method. amx_bf16, amx_tile, amx_int8, avx512_bf16; Xeon @ 2.10 GHz, 4 vCPU; tier AMX TDPBF16PS; main bc8012da. cargo test --release --test tile_gemm passes (5 tests). Five runs of tile_gemm_cost, each the example's own medians of 7.

Map	popcount counts	AMX A·A	stage	bit compose	bit triangles	BF16 triangles
1 %	181 (164)	88 (67)	182 (132)	2.1 (2.0)	2.5 (1.9)	57 (45)
10 %	184 (135)	80 (58)	222 (182)	7.6 (6.0)	10.8 (7.4)	110 (94)
50 %	174 (137)	84 (63)	501 (470)	34 (27)	48 (36)	422 (323)
hubs	174 (128)	82 (58)	171 (143)	3.7 (2.6)	6.0 (3.7)	63 (41)

Median of five, in µs; #1626's value in brackets.

What holds:

• For walk counts, AMX is faster than popcount by 2.06–2.31×; D-RMAP-2 §5a-AMX: the tile arm measured on the AMX tier #1626 measured 2.2–2.5×.
• Staging pays for itself after 1.85–5.57 products, against D-RMAP-2 §5a-AMX: the tile arm measured on the AMX tier #1626's 1.4–6.4.
• The bit arm still wins Boolean composition by 6.6–71× and triangles by 8.8–22.6×.
• The threshold alone costs more than the bit composition on every map.

What differs:

• Absolute times are 10–60 % higher in every cell. This looks like load on the shared VM, not a code change, but I have not verified that.
• The row-major entry took 7.8–9.0 ms against D-RMAP-2 §5a-AMX: the tile arm measured on the AMX tier #1626's 5.6–6.1 ms. That entry doesn't use AMX, and why it is faster than §5a's 19 ms remains open.
• bf16_bytes now reads 393 216, so D-RMAP-2: count the row-major B blocks in the BF16 byte total #1625's fix is on main.

Arithmetic check. Every derived claim in #1626's body also matches its own table: 2.18–2.46×, repaid after 1.36–6.35 products, composition 6.1–59.6×, triangles 8.9–23.8×.

Gates

• entries_index.py --check: current (383 rows).
• append_only_gate.py: no protected file shrank. All five prepended files grew: PR_ARC_INVENTORY +42, LATEST_STATE +7, ISSUES +53, INTEGRATION_PLANS +4, AGENT_LOG +9.
• supersession_index.py: regenerates unchanged.

---

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: acaf48cc-5d5b-4c94-82da-a9728429e209

• 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 2026-10-20T10:17:02.849075Z	a2286d2	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.