Housekeeping: PR triage, doc staleness, and alignment with current research direction

[0bserver07]

Summary

End-of-month housekeeping: 3 open PRs need decisions, several CLAUDE.md /
docs/ surfaces are stale relative to current ByteDMD work, and the
experiments/grid machinery in cybertronai/ByteDMD@dev introduces concepts
this repo's docs should reference. Single issue for trackability; will split
into focused issues once direction is confirmed.

Open PRs — recommendations

PR	State	Recommendation
#94 — KM-min + SAT hybrid (@SethTS)	LGTM from @yaroslavvb; author replied with rationale	Merge. Pre-merge nit: body references a local-only path (~/dev/research/sutro/km+smt.md) — strip from squash message or move into repo.
#87 — ByteDMD floor-gap survey (@SethTS)	Commits shipped via squash of #88; diff vs main is empty	Close as already-shipped. Two substantive open threads (oracle-query KM fairness; "is km-min actually a lower bound?") deserve dedicated issues.
#63 — bin/review-cycle cross-model supervisor (@philoengineer)	Open since Mar 28, no reviews	Decide this week. Concept is sound but predates the ByteDMD shift. Either review and merge, or close politely with reasoning.

Doc staleness — needs an update pass

• CLAUDE.md "Current Best Methods" table is already flagged as pre-ByteDMD
  legacy DMC numbers. Beyond unit drift, the framing ("lowest ByteDMD score
  wins") is now downstream of the broader cost-model work in
  cybertronai/ByteDMD@dev and needs revisiting alongside the table.
• docs/research/bytedmd.md does not link to the Manhattan-diamond model,
  the sandwich theorem (0.3849 · bytedmd_live ≤ manual ≤ 4.0 · bytedmd_live),
  or the lower-bound family (global_density, local_density,
  polymatroid_lb, bytedmd_opt).
• Telegram local DB is missing 3 of 9 forum topics (challenge Update TODO.md: mark Phase 1 tasks complete #2,
  Pitch / Talking Points, makemore task results). Sync only picks up topics
  it has seen — needs a small bootstrap.
• docs/catchups/ had no entry between Mar 22 and Apr 30. The
  2026-04-30 catch-up just landed; future cadence should be weekly.

Reading queue (context, not action items)

Two Google Docs were posted late-April that frame the current direction.
They are not yet in the docs/google-docs/ manifest:

• Bounds summary (Apr 28):
  https://docs.google.com/document/d/1TcRBets7L1TLF628uUe8DOoNcGGxM7rl5ECZ4mT25ks/edit
• Higher-level thinking (Apr 30):
  https://docs.google.com/document/d/17eexzDhYJGl2FJIsdacL8ikMzNSBV38Ji01MK6YHqWI/edit

Public ByteDMD references worth surfacing in this repo

Active research front in cybertronai/ByteDMD@dev has new public artifacts
this repo's docs do not yet point to:

• docs/manhattan-diamond.md
  — simplified explicit communication model: 2D Manhattan grid, ⌈√addr⌉
  per read, writes/arithmetic free.
• docs/instruction_set.md
  — complete list of tracked Python ops and arities.
• experiments/grid/README.md
  — 45 algorithms × 7 placement strategies / bounds: hand-placed manual
  schedules, three LP lower bounds (global_density, local_density,
  polymatroid_lb), Bélády MIN, and the LRU live/classic variants.
• experiments/grid/manual_dsl.py
  — DSL for hand-placed schedules; relevant to any future
  "submit an IR program" challenge interface.
• experiments/grid/grid.md
  — measured numbers across the 7 columns.

Two relevant open ByteDMD issues (cybertronai/ByteDMD#2, #3) ask whether
loops are unrolled and whether memory management is explicit — worth tracking
from this side.

Proposed follow-up issues (not yet filed)

If the direction below makes sense, will file as separate issues:

1. Compute polymatroid_lb, global_density, local_density for sparse
   parity — directly addresses the "is km-min actually a lower bound?"
   question. Reuses ByteDMD's existing LP machinery.
2. Adopt manual-style placements as a new submission category for
   cybertronai/sparse-parity-challenge — the current solve() interface
   traces numpy and is mismatched with hand-placed schedule evaluation.
3. Mirror Manhattan-diamond cost model into this repo's docs — short
   page summarizing the 2D grid, the sandwich theorem, and how
   bytedmd_live relates to the LP bounds. ByteDMD repo as source of truth.
4. Bootstrap fix for Telegram topic sync — pick up newly-created
   forum topics on next sync rather than requiring a manual seed.
5. Add the two late-April Google Docs to src/sync_google_docs.py
   manifest.

What I'll do this week, pending sign-off

• Triage actions on PRs exp: KM-min + SAT hybrid — first ByteDMD measurement of SAT backtracking #94, exp: ByteDMD floor-gap survey — KM-min 268 vs GF(2) 101,501 #87, Add cross-model research supervisor (bin/review-cycle) #63 per recommendations above
• File the 5 follow-up issues above if confirmed worth doing
• Additive doc PR: extend docs/research/bytedmd.md to reference the
  Manhattan-diamond model and the sandwich theorem (no tracker code
  changes)
• Add the two new Google Docs to the sync manifest
• Re-establish weekly catch-up cadence (next one ~May 7)

Will not touch experiment code, locked measurement files, or the
"Current Best Methods" table re-measurement until the framing is confirmed.

---

agent-0bserver07 (Claude Code) on behalf of Yad

[yaroslavvb]

We're looking at https://github.com/cybertronai/sutro-problems right now, the whole bound stuff ended up kind of complicated, although there were useful learnings maybe we'll reuse later for the 2D grid stuff