Repository navigation
PP-TENSOR-001 T1: typed TensorStorage — a dense consumer cannot receive an MoE placeholder #3433
Description
Parent: #3428 · Slot: 0.69.0 must-carry, directly after M1.
Mechanism: dense_ffn_placeholder is an ordinary tensor ref with 0 bytes; the contract "check moe_layers[i].is_some() first" lives in a doc comment, and #2535 was a consumer two files away that did not read it.
Fix: absence is a variant, not a value; dense FFN fields on a MoE layer are Absent. The compiler then lists every consumer.
Falsifier: a dense forward over a MoE layer fails to compile, or returns one typed error at one site; M3 fixtures for #1749/#1789/#2535 stay GREEN with the three consumer-side guards deleted.
Ĵ: 24 turns [A] — borrowed; estimate.sh aprender pools 0 of 47 ledger rows today, so there is no measured basis.
Plan + evidence: #3427 · design discussion: #3418 · spec PR: #3420
🤖 Generated with Claude Code
Launch conditions (operator decision 2026-09-17): starts after #3430 merges. PLAN PHASE FIRST — the plan quorum rules on T1 here (and T2 for #3434) using Table T from the PMAT-3427 receipt: 5 carriers (QuantizedTensorRef, OwnedQuantizedTensor, TensorInfo, RawTensor, Qwen3MoeQuantizedLayer), every one can hold (len = 0, qtype = 0), zero validating constructors; 29 and 26 literal constructor sites for the first two. First task inside the ticket: replace Table T's file-level "unguarded" regex with a call-site count via pmat query — the file-level figure (63 files read the dense FFN fields, 14 with no guard token) is weak evidence and is not the migration size.
Activity
noahgift commentedon Sep 23, 2026
milestone → 0.74.0 per docs/roadmaps/milestone-triage-2026-09-23.md (#4024), operator approval 2026-09-23 (relayed by aprender-cf). Reason: quant/tensor/arch dispatch consolidation: Any Model core (#4001)
noahgift commentedon Sep 25, 2026
Parked by APR-EPIC-001: over E7 budget (rank 5/12). Reopen = pull. Applied by aprender-d1: claude+agy both said close, apr shadow row recorded (handoff/close-quorum-2026-09-25.json).
docs(lookahead): A3 measured, #3433 closed not-planned and the 0-byte…
noahgift commentedon Oct 15, 2026
Reopened (pull) for the 0.74 Any Model plan, row A3, at the request of the session that will do the work (aprender-w71-e3a, branch off main 8f601ca, deadline 07:30Z). The NOT_PLANNED close (2026-09-25) was a parking decision made before the 0.74 plan, and the park comment says reopen = pull. The falsifier is live on main: the 0-byte dense_ffn_placeholder in the GGUF transformer loader (offset 0, byte_size 0) and the skipped MoE CUDA parity gate. Milestone 0.74.0 kept. Refs #4001.
noahgift commentedon Oct 22, 2026
Tracked in #4001 (triage 2026-10-22).