mask-risc: collapse ≤3-leaf Boolean op chains onto the ternlog Count/Any fold

[AdaWorldAPI]

Summary

Before this PR, the Count/Any fold from #1270 only handled a single Boolean op. Now any raw op chain over at most three resident planes gets the same treatment: Program::fused_ternlog() interprets the whole chain symbolically.

• How it works: each plane is a leaf truth table (0xF0 / 0xCC / 0xAA). Each op combines tables using its own Boolean law; Ternlog applies its immediate bitwise to the three input tables. Each op writes the resulting table to its destination slot, so slot reuse is modelled exactly.
• What the fold receives: the terminal slot's table becomes the fold immediate.
• No allocation: recognition uses a fixed on-stack array of FUSED_SLOT_CAP slots, the same bound the validation bitmap uses.

Stays on the tiled path, unchanged:

• a fourth distinct plane;
• a Pred (comparison leaf) or Gather anywhere in the chain;
• a scratch slot read before it is written;
• a slot at or above the cap;
• any terminal other than Count/Any (Keep remains the explicit election of a bitmap).

Scope: lane predicates are deliberately not attempted. Chains with more than three leaves are measured at their boundary only. No Rayon, OQ-5 or scheduler changes.

Gates

• End to end: tests/program_collapse.rs checks that fuse_program output reaches the fold.
• Differential: raw chains are checked against the tiled Keep path across extents, and against reference_execute over the whole population. The chains are a catalogue of 7 (slot reuse in place, a derived-input ternlog, a dead op, double complement, a & !a) plus 400 random chains, with an anti-vacuity check of at least 20 distinct counts.
• Zero materialization: a poisoned-scratch gate plus a counting allocator, each with a Keep twin as a control.
• Recogniser test updated: in tests/fused_ternlog.rs it now admits exactly the collapsible chains. Two exec.rs fixtures that needed an unfused program now use Terminal::All, because Not and chains fold now.
• Disable runs: committed first, confirmed red, then restored to green. Each of these fails inside program_collapse.rs:
  • an interpreter limited to single ops;
  • a transposed table-apply;
  • slot reuse ignored;
  • a fourth leaf aliased to leaf 0.
• Checks: full lance-graph-mask-risc and lance-graph-quack suites pass; clippy -D warnings and fmt are clean.

Measurement

cargo run --release -p lance-graph-mask-risc --example program_collapse_probe. Setup:

• Tier: x86-64-v3 (the AVX2 polyfill, the repo default). v4 was not measured.
• Size: N = 1M rows.
• Values: median, oracle-checked.

chain	ops	fold Count	tiled Count	derived words written by the tiled path
(a&b)|!c	3	8.6 µs	224 µs	49 152
((a^b)&!c)|(a&c)	4	16.2 µs	352 µs	65 536
maj(a,b,c)^a	2	17.4 µs	166 µs	32 768
a&d, empty by data (Any: 4.4 vs 111 µs)	1	6.5 µs	120 µs	16 384
(a&b)|(c&d), 4 planes	3	not collapsed	283 µs	49 152

How to read the numbers:

• The fold writes 0 derived words.
• Any on non-empty data: the gap is dominated by the fold exiting at its first block. The data-empty row isolates the scan itself: 25× for Any, 18× for Count.
• The gap is not all write elimination: the tiled path also pays per-tile interpreter dispatch, and this probe does not separate the two costs.
• An algebraically constant chain (e.g. a&b&!a → table 0x00) monomorphizes to a fold that never reads the planes. That is why the empty row is empty by data.

Board entry: .claude/board/entries/2026-09-23-program-collapse-boolean-chains.md

🤖 Generated with Claude Code

https://claude.ai/code/session_01GXUahz73MZxtxWcfpHp9dG

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Essentials

Run ID: 44ebbf2f-962d-4cad-9ced-74d7ee005f12

📥 Commits

Reviewing files that changed from the base of the PR and between d90600d and c47698f.

📒 Files selected for processing (7)

• .claude/board/entries/2026-09-23-program-collapse-boolean-chains.md
• .claude/board/entries/README.md
• crates/lance-graph-mask-risc/examples/program_collapse_probe.rs
• crates/lance-graph-mask-risc/src/exec.rs
• crates/lance-graph-mask-risc/src/ir.rs
• crates/lance-graph-mask-risc/tests/fused_ternlog.rs
• crates/lance-graph-mask-risc/tests/program_collapse.rs

 _________________________________________________
< Because, even your code needs a second opinion. >
 -------------------------------------------------
  \
   \   \
        \ /\
        ( )
      .( o ).

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

---

_{Comment @coderabbitai help to get the list of available commands.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_a21e4a07-c99c-44da-b2a1-4675707486e6)