You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
splice: fn-local consts crossing module boundaries (train_hslm MAT_V/MAT_DOWN undeclared after the var pull) #13067
The use_resolve var splice (decl_name learning var, + the collision suppression) unblocked wasm_doom_fb_link but breaks generation for specs that pull cross-module fn-local consts: t27c test specs/ml/train_hslm.t27 emits broken zig -- use of undeclared identifier 'MAT_V' / MAT_DOWN (t27_test.zig). Facts measured on the PR tree:
MAT_V/MAT_DOWN (and MAT_Q/K/O/UP/OUT/NONE) are declared fn-local (indent > top) in specs/nn/hslm.t27 (~lines 110-119), so split_decls does not see them as declarations and available[] never keys them; yet at the merge base train_hslm generated clean -- the base splice carried them some other way, and the var pull changes which decls exist/break that path.
With the splice, t27c typecheck of train_hslm passes (0 errors; was 56 under the keyword-only commit), so the break is generation-stage only, and t27c seal --save refuses (tests fail), producing gen-drift in seal coverage.
Seal coverage at the base commit is green (rc=0 locally): the regression is ours, not master drift.
Remedy direction: the splice's frontier must resolve identifiers that live inside pulled decl bodies (fn-local consts of the used module) the way the base path did, or decl extraction must key them. Until then the var splice cannot land on master; it stays on the branch lane. Handoff from doom-ci-repair, 2026-11-01. Refs #13059#12387.
Activity
gHashTag commentedon Nov 17, 2026
Spec created: specs/tri/splice.t27 (typecheck OK, 1 test passes).