profile: execute the MoE region terminator, and stop shared work inheriting pair labels

[qichao-arlo-wang]

Companion to AICrossSim/PLENA_Compiler#72; the submodule pin moves onto it here.

Land #72 first. The vocabulary guard holds StageKind equal to MOE_STAGES in both directions, so neither side can merge alone. With #72 pinned and this branch's StageKind reverted, the guard fails with compiler emits @stage=non_moe but no StageKind matches it — that is the lockstep doing its job, not a problem.

1. The MoE region now has an end

StageKind::NonMoe resolves the compiler's @stage=non_moe. Demonstrated end to end — a shared-expert program followed by an lm_head projection, classified with the emulator's own sticky-marker rule:

=== WITHOUT terminator ===  219 instructions
    shared_expert_projection       116   53.0%   <-- billed to a MoE stage
    shared_expert_activation       103   47.0%   <-- billed to a MoE stage

=== WITH terminator ===  219 instructions
    shared_expert_activation       103   47.0%
    shared_expert_projection        87   39.7%
    non_moe                         29   13.2%

29 instructions of pure non-MoE work stop being billed to the shared branch, and that number feeds shared_branch_fraction.

A distinct variant, not other. Resolving other would have worked mechanically, but a large Other is a coverage problem and a large NonMoe is just a long epilogue. Being unable to tell those apart is how the original bug stayed invisible.

2. Shared work no longer inherits a pair label

The three shared_expert_* stages have no experts — there is no (token, expert) pair for their work to belong to. Without a reset their cost merged into whichever routed pair ran last, and in a shared-only program it invented pairs outright.

The reset list is now PAIRLESS_STAGES, a named constant, so the test asserts membership instead of restating the match arm — the two could previously drift.

3. An unterminated region is now visible in the JSON

classification.trailing_stage and classification.trailing_instructions.

The per-stage totals still add up when a region is left open — that is precisely what made this class of bug invisible. A large trailing count against a MoE stage now says so, without anyone having to read the source that emitted it. A properly terminated program shows non_moe here and the count is simply how long the epilogue is.

On test_branch_split.py

non_moe is classified explicitly, in its own NON_MOE_STAGES tuple rather than folded into PLUMBING_STAGES. Plumbing is MoE cost that both branches share; this is not MoE cost at all. The exhaustiveness assertion is what forced the decision rather than letting a new stage drop out of the split and read as free — working as intended.

Validation

• cargo test --workspace --release — 181 passed, 0 failed (3 new: terminator resolution, trailing attribution both ways, and pairless-stage membership including the negative case that a routed stage keeps its label).
• Cross-repo vocabulary guard passes with the new pin.
• testbench guards 10 passed; just test-moe-shared-all passes with 3 bit-exact comparisons; all six routed-MoE suites pass.
• cargo fmt --check, ruff format --check, ruff check clean.

🤖 Generated with Claude Code