suggestions.md

Runtime — suggestions

CT-derived only. Each cites the rule it applies.

1. Work should not carry an Arc<Run> — §5 deduce, don't store

Every WorkItem clones an Arc, so a dispatch of S slices pays 2S atomic refcount operations to re-state a fact the run already knows: the lane is executing this run. The run outlives every item it queued (mapal_par_finish is help_until(remaining == 0)), so the lifetime is already implied by the protocol and the refcount is a stored copy of a deduced morphism.

2. take_any's miss path is the last O(N) — §4.5 Law 4

A failed steal locks every lane's Mutex<VecDeque> in turn. There is no shared record of which lanes hold work, so each thread rediscovers it by polling all of them — FRAMEWORK §5 "one source of truth for shared structure" failing on the idle set. Pool::pending (added S50 P1) is that record for the spin loop; the steal path does not use it yet.

3. The ARCHITECTURE §4 composition rules want pins — §6.6

Rule 1 (unlock before publish) was violated for the whole of S50 P2 and surfaced only because P1 made the host fast enough to observe it. A rule that can only be caught by an unrelated performance change is an undocumented exception waiting to rot. Each of the six is testable from mapal-rt's own suite.

4. path_plan_is_acyclic should become unnecessary — §6.6

llvm/src/lib.rs:258 carries a ponytail: note saying to delete the fallback "when path_plan's total-DAG contract is enforced before backend entry". S49 removed the sibling-merge cycle class but not the general one, and a cyclic plan is still silently demoted to sequential emission — the failure mode that hid every chained matmul for nine sessions. Either enforce the contract or make the demotion loud.

5. The dispatch verdict is compile-time; the cost that decides it is not — §4.2

DispatchCost::dispatch_ns is a property of the runtime, measured on a program, and it is consumed by the emitter. That is a TrnLoc fact crossing into a compile-time decision with no mechanism keeping the two in agreement — exactly the drift shape §6.3 exists to prevent. A runtime self-calibration at pool creation, published where the next build can read it, would make it a measurement instead of a written-down number.