Repository navigation
EPIC: PP-QUANT-001 — quant-type dispatch, Phase 0/1 (0.69 train) #3421
Description
Tracking epic for #3418's quant-type dispatch consolidation — scoped to 0.69,
Phase 0/1 only. Design/phase plan: docs/specifications/PP-QUANT-001-MASTER.md.
Phase 2/3 tracked separately by epic #3423 (dedicated milestone, not this train) —
see the placement note below.
Placement decision (2026-09-17): Phase 0/1 rides the 0.69 train, overriding
#3418's own request for a dedicated non-cadence release, because both are additive
and touch no live call site. Phase 2 (the ~30-file call-site migration across
CPU/CUDA/wgpu) and Phase 3 (Qwen2.5-Coder parity re-proof) are the large-blast-radius
part #3418 actually asked to be dedicated — those now live on milestone
Inference dispatch & architecture consolidation, tracked by epic #3423, alongside
PP-ARCH-001 (#3422). This epic closes when Phase 0/1 does.
Phases (spec §3) — THIS epic
- Phase 0 — Inventory:
contracts/quant-dispatch-completeness-v1.yaml(PR spec(PP-QUANT-001): quant-type dispatch consolidation — design, contract, Phase 0 gate #3420) - Phase 1 —
QuantCodebookFormattrait, extendedQuantBlockFormatregistry,quant_type_traits()dispatch fn, remainingqdc_*tests
Phases tracked by #3423 instead
- Phase 2 — Migrate all ~30 call sites off local match arms onto the dispatch fn
- Phase 3 — Full parity re-run (Qwen2.5-Coder GPU-beat) vs. pre-migration baseline
Owner: unassigned.
Activity
alfredodeza commentedon Sep 17, 2026
Phase 0 — Inventory: done in PR #3420 (pending review/merge).
contracts/quant-dispatch-completeness-v1.yamlis the single enumeration: 35 live ggml tensor types (ids 0–42, 8 removed,type_count43 — verified against llama.cpp3173a5647), geometry fromggml-common.hstatic_asserts, ten codebook grids, and the 30 dispatch sites.qdc_dispatch_sites_match_oraclefails CI on any new raw-qtypematch site or any row flipped tomigrated: truebefore the site actually stopped matching.Phase 1 design is in the spec (§2.4) — the table is built on
GgmlQuantType, entries are plain data +fnpointers (not trait objects;QuantBlockFormatis not object-safe), hot kernels stay monomorphized. Owed for Phase 1:dispatch.rs,codebook_trait.rs, generatediq_grids.rs+ its generator, the row-regeneration script, and the two remaining tests (PO-QDC-001/004).docs(spec): PP-QUANT-001 — cross-link downstream PP-ARCH-001, close o…
[-]EPIC: PP-QUANT-001 — quant-type dispatch consolidation (0.69, phased)[/-][+]EPIC: PP-QUANT-001 — quant-type dispatch, Phase 0/1 (0.69 train)[/+]docs(spec): PP-QUANT-001 — Phase 2/3 gets its real milestone, not "ma…
noahgift commentedon Sep 17, 2026
Decisions for a human — one comment, from the PMAT-3427 research run (2026-09-17). Nothing of yours was moved, re-titled or re-scoped.
1. Placement conflict. #3423 puts PP-QUANT Phase 2/3 on the dedicated non-cadence milestone. The operator's session brief of 2026-09-17 says: "Alfredo's ticket is prioritized: PP-QUANT-001 is 0.69 must-carry" — a slice M that rides 0.69.0, with T-0 waiting for it (andon at 2026-09-20T06:35Z, 72 h after the actual
v0.68.0tag). Slice M is deliberately smaller than Phase 2: #3429 fixtures → #3430 oneGgmlType+TRAITS[43]below core/compute/serve → #3431 reconciliation gate (builds on PR #3420'sdispatch_contract_tests.rs) → #3432tensor_byte_size+ single refusal site (fixes the #3091 reopen for all 35 ids, no new kernel). The per-directory retype tickets #3436–#3440 restate your Phase 2; they sit in 0.70.0 only as a placeholder and should fold into whichever placement stands. Needed: confirm slice M on 0.69.0 alongside #3423 owning the migration, or say otherwise.2. A second defect class neither spec covers. 3 of the 5 crashes cited in #3418 are MoE placeholder tensors (
byte_size: 0,qtype: 0= F32) reaching dense consumers; all three fixes are guards, the placeholder is still built onmain. Proposed as sibling epic #3428 (PP-TENSOR-001; T1 #3433, T2 #3434 on 0.69.0; T3 #3435 on 0.70.0). Evidence table in the PR #3420 review. Needed: accept as a sibling under #3418's umbrella, or fold into PP-ARCH-001.3. #3091 is open in milestone
0.68.0, which shipped (open 2 / closed 93). Reproduced on the published 0.68.0 asset with a 338 MB public model. Human-authored, so not moved. Needed: move to 0.69.0 (its fix is #3432). Process ticket: #3445.4. 0.69.0 cannot reach a clean T-0 as it stands
[V]: 124 open one day before its due date — 109 with no linked PR, 1 with an open PR (#3204), 1 whose linked PR merged with the issue still open (#2753 → #2809), 3 epics, 10 minted today. Authors: noahgift 118, alfredodeza 5, guyernest 1. Every row'supdatedAtis within 7 days, so that field measures triage edits, not work. Needed: an owner-level bulk slip of the no-PR rows to 0.70.0; this run moved only tickets it minted.Full tables:
docs/audits/impl-PMAT-3427-receipt.mdin the docs PR for #3427.noahgift commentedon Sep 17, 2026
Mapping — my per-directory retype tickets are closed into #3423 Phase 2 (operator decision 2026-09-17)
Alfredo's tickets win on duplicates, so #3436–#3440 are closed, the same way #3444 was closed into #3422. What each one measured, so nothing is lost when Phase 2 is planned. Counts are from a file:line scan of the 33-file union (PR #3420's
dispatch_sites∪ every non-test file naming aGGUF_TYPE_*constant), onorigin/main@7eb81a531[V]:gguf/loaderstensor_byte_size)gguf/inferenceCPUacceleration.rs13 ids,matmul*.rs10,forward/*3 — a tensor accepted by one path is refused by its neighbour;matmul_fused.rsandfused_matmul_into.rshold 61 of the 231 sitesgguf/cuda+cuda/WeightQuantType(8) is a fourth registrygpu/adapterswgpuGgmlType12 ids (import path), compute 15; 38Quant*/Ggml*enums in the workspace, 13 of them naming ≥ 5 ggml typesaprender-serveonly; the cross-crate part stays with #3430 (one enum below all three)Scan summary: 231 code sites in 28 of the 33 files — 151 arm/reference, 54
.qtype ==/!=comparisons, 24match, 2 bare-integer (matches!(qtype, 12..=14)inapr/cache.rs). The other 5 files of the union are not sites: 2 are whole-file test modules without "test" in their name (gguf/model_owned_quantized.rs,gguf/transformer_quantized_layer_field.rs) and 3 name a constant only in a comment or aconstdefinition. That also corrects my earlier "27 files": 5 of them are not code sites, so it is 22 real + 6 found only by the spec's oracle = 28.Still open under this epic from my run: slice M — #3429 → #3430 → #3431 → #3432 (0.69.0 must-carry, confirmed by the operator today), then #3441 tier-0 dequant, #3442 drift job, #3443 generated table (0.70.0).
noahgift commentedon Sep 17, 2026
0.69.0 — per-row verdict (124 open rows, read 2026-09-17). Verdicts only: nothing human-authored was moved.
Rule (operator, 2026-09-17): stay = must-carry, or a linked PR that can be green before the andon at 2026-09-20T06:35Z; otherwise remove the milestone (not → 0.70.0). Mechanical inputs:
closedByPullRequestsReferencesand each PR's check rollup[V]. A PR that mentions an issue without a closing keyword is invisible to this query — rows you know are in flight are yours to overrule.Counts: remove milestone 110 · stay, at risk 2 · close-check 1 · stay 11. Authors: noahgift 118, alfredodeza 5, guyernest 1.
#[repr(u32)] GgmlType(43 rows: 35 live + 8 Removed) + `TGgmlTypevs the vendoredggml.hid litensor_byte_sizeand ONE refusal site readTRAITS— closesTensorStorage— a dense consumer cannot receive an MoEqtype: 0+byte_size: 0may not be construcmain— land §1.5 (d76b4a417, CI pending; stays draft until a quorum reviews itapr run: 7 defects incl.--streamemitting empty text and--trace-outputaapr servehas three different HTTP surfaces and which one you get dependsKaniStrategyhas no label for symbolic-f32-array harnesses, the dominant DSPPOST /api/generateandPOST /generateenforce opposite unknown-field policieignore_eosandseed: apr-cli's ChatCompletionRnoahgift commentedon Sep 17, 2026
Quorum status: did not start —
STOP(rail), 2026-09-17. The three-lane review of Q1 (oneGgmlType+TRAITS[43]), T1 (typedTensorStorage), T2 (sentinel ban) and X1 (ordering) cannot be dispatched under the plan ticket: the harness refuses it before any lane —kind-gate: kind=docs is not implemented … nothing was planned, dispatched, or written for PMAT-3427(exit 3),model-gate: unsupported kind 'docs'(exit 2). I did not relabel the ticket or substitute my own review for a lane. So there is no verdict grid, and #3430, #3433, #3431, #3432, #3434 stay unstarted per the operator's instruction; #3429 and #3445 have no design dependency. The andon check at 2026-09-18T12:00Z stands: unless the harness gains akind=docsroute (or the quorum is run under a different ticket by the operator's decision) it will not have returned by then.One measured input the quorum would have used, for whoever runs it: X1's falsifier is "carrier structs and raw-integer qtype sites overlap < 20 % ⇒ T1 and Q2 can run in parallel". Measured on
7eb81a531: 14 of 34 serve raw-integer sites (41 %) sit in a file that uses a tensor carrier — the falsifier does not fire, T1-before-Q2 stands on the evidence. Tables are in PR #3447's receipt.noahgift commentedon Sep 17, 2026
Where the design review happens now (operator decision, 2026-09-17)
The plan ticket (PMAT-3427 / #3427) cannot host a quorum — the harness refuses
kind=docs. Rather than relabel or bypass, the design quorum moves into each code ticket's own plan phase (kind=codeis implemented). PMAT-3427 stays as is; draft PR #3447 stays draft until the harness gains akind=docsroute, then gets its own quorum.GgmlType+TRAITS[43])mergeorkeep(concept); home crate forGgmlTypederived fromcargo tree. No code until 3/3. Verdict grid gets posted on PR #3420.TensorStorage)(len=0, qtype=0), zero validating constructors. First task: a call-site count viapmat queryreplacing the file-level "unguarded" regex.[U]today) — the gate needs that as its baselineX1 stands on measurement: 41 % of serve's raw-integer qtype sites sit in a file that uses a tensor carrier (file-level; falsifier was < 20 %) ⇒ T1 before Q2. Carried numbers are now in #3430's body: 231 dispatch sites in 28 files, 53 raw-integer fields/params.
This supersedes my earlier "quorum did not start" comment: the 2026-09-18T12:00Z check no longer applies to PMAT-3427; each code ticket's plan quorum has its own clock. Slice M's andon is unchanged — 2026-09-20T06:35Z.
noahgift commentedon Sep 17, 2026
Start now — do not wait for #3448 to merge
The roadmap entries for these tickets were stuck in a draft PR; #3448 (entries only, auto-merge armed) puts them on
main. Nobody needs to wait for it.kind-gatereads the roadmap in your working tree, so in a fresh branch offmain:Measured on a clean
origin/mainworktree for PMAT-3430:kind-gateexit 2 → exit 0. The fragment is byte-identical to the one in #3448, so it merges clean when #3448 lands; ifroadmap.yamlconflicts, regenerate it with the same command.Which model can run which ticket (
model-gate, measured): PMAT-3430 and PMAT-3433 now carryorch:fable+orch-basis:M>=3(core + compute + serve) — a Fable session is admitted. PMAT-3429 and PMAT-3445 have no basis for Fable and are Opus-class: launch those with/model opus, or a Fable session is refused at Phase 0.noahgift commentedon Sep 17, 2026
Milestone triage — REVERT CSV (written before any milestone move)
Snapshot 2026-09-17,
gh issue list/gh pr list --state open: 382 issues + 31 PRs = 413 rows. Emptyold_milestone= none. Operator-authorized triage run; this is the first write of the run. To revert one row:gh issue edit <n> --milestone "<old>"(or--remove-milestonewhen empty).revert.csv (413 rows)
noahgift commentedon Sep 17, 2026
Draft PR with the plan and receipt: #3452
PMAT-3430 plan quorum — history, v1–v4. Not 3/3. No v5 written; no code written.
Plan:
docs/audits/impl-PMAT-3430-plan.mdon branchPMAT-3430-one-ggml-type(draft PR linked below). Lanes: agygrillme, review-only, none in the author's model family (authorfable-5-1). Every objection below was re-executed by the orchestrator before it was answered; two lane claims were refuted on re-run (lane 1 v1: "assoc-const alias works in patterns" — rustc E0308; lane 1 v1: "Q8_1 = 40, C struct has padding" — upstreamstatic_assertsays 36).1. Grid
† lane exited 3,
LANE ISOLATION VIOLATED — a shared ref changed: the plan worktree shares a git dir with the main checkout, and other sessions moved refs during the lane window (v2:refs/remotes/origin/main; v4:refs/heads/PMAT-3445-t0-milestone-gate,refs/heads/PMAT-3451-ont-6-verdict-lattice, a merge-queue ref). Each lane's own clone was unaffected, but under the rail those verdicts are unclean — v3 is the only round with three clean lanes, and it was 1/3. v1'spartialwas 2 stderr lines per lane beyond workspace narration.2. Every REJECT / AMEND reason, verbatim from the lane's
findings[].claim, one tag eachcrates/aprender-compute/src/inference/model.rs:712docs/audits/impl-PMAT-3430-plan.md:58crates/aprender-serve/src/paged_kv/mod_compute_prefix.rs:254crates/aprender-core/src/format/tensors_tests_workflow.rs:82crates/aprender-serve/src/apr/mod.rs:324docs/audits/impl-PMAT-3430-plan.md:56docs/audits/impl-PMAT-3430-plan.md:57docs/audits/impl-PMAT-3430-plan.md:46docs/audits/impl-PMAT-3430-plan.md:49docs/audits/impl-PMAT-3430-plan.md:101docs/audits/impl-PMAT-3430-plan.md:101docs/audits/impl-PMAT-3430-plan.md:76docs/audits/impl-PMAT-3430-plan.md:31docs/audits/impl-PMAT-3430-plan.md:56crates/aprender-serve/src/gguf/tests/loader_tests_apr.rs:311crates/aprender-serve/src/apr/mod.rs:324docs/audits/impl-PMAT-3430-plan.md:99cargo test -p aprender-quant --lib ggml_typedoes not run integration tests in crates/aprender-quant/tests/ (declared in Phase 1 scope_paths).docs/audits/impl-PMAT-3430-plan.md:101cargo test -p aprender-serve --lib gguf::typesruns only 10 unit tests in types.rs and does not execute qwen3_moe_load.rs (#3405 QTYPE_LABELS removal) or loader_tests_apr.rs.docs/audits/impl-PMAT-3430-plan.md:59core: format/gguf/reader.rsis an id-to-enum boundary and will get an admission function.crates/aprender-core/src/format/gguf/types.rs:361GgmlTypein core that will become non-exhaustive.crates/aprender-core/src/format/gguf/types.rs:361docs/audits/impl-PMAT-3430-plan.md:45crates/aprender-core/src/format/gguf/reader.rs:325crates/aprender-serve/src/gguf/qwen3_moe_load.rs:107docs/audits/impl-PMAT-3430-plan.md:45docs/audits/impl-PMAT-3430-plan.md:40docs/audits/impl-PMAT-3430-plan.md:59docs/audits/impl-PMAT-3430-plan.md:69docs/audits/impl-PMAT-3430-plan.md:110aprender-quantwithout doc comments will fail the crate's existing#![warn(missing_docs)]lint, though it does passnon_camel_case_types.crates/aprender-serve/src/apr/mod.rs:324Makefile:1129crates/aprender-serve/src/apr/mod.rs:324crates/aprender-serve/src/infer/mod.rs:38contracts/census.json:5README.md:44ci/explicit-test-commands.d:1.github/workflows/ci.yml:938Counts by tag: Q1 (enum + TRAITS design) 21 · plan mechanics 16 · Table E resolutions 2 · home crate 0 · X1 ordering 0.
aprender-quantas home was called sound by every lane in every round; no lane contradicted X1; Table E drew one flip (KvQuantType, lane 1, v1 only) and one path-format note.3. Objections that repeated across rounds
No objection came back in the same words — each round's specific was re-run, answered, and not raised again. But three objection classes survived two or more rounds, each time with a new instance the previous revision had not found:
from_id → Resultandtensor_bytes → Optionbreak callers · v2: block omitsas_byte as_str from_str_lossy block_bytes block_size Display· v3: aux types lack derives,const fncalls non-const,must_use/missing_docsunder the crate's lint tableapr/mod.rs:324+ 5 more bypass it · v2: listed corereader.rs, which keepsu32raw — core has no boundary · v4:apr/mod.rs:324is an inline match insideTensorEntry::from_binary, not a callable "boundary function that keeps its name"format/siblings, serveapr/infer/;gguf::typesfilter exercises 0 sites · v3: Phase 1 scope omits servegguf/dtype.rs; tests on privatefrom_u32break in Phase 3 · v4:contracts/census.json,README.md:44(1791→1792),ci/explicit-test-commands.d/, MakefileCONTRACTSare outside every phase's scope; "CI selection line" does not exist (guards auto-dispatch viaguard_tree.sh --no-cargo)Read against the operator's rule: A and B are tagged Q1 and each survived three rounds — by that rule they are design findings, not wording. What they have in common: the plan keeps trying to make M1 source-compatible and behaviour-preserving at every existing call site while replacing three enums with one 35-variant enum, and each round finds one more site where that promise needs another special case (an alias, an
Option-returning twin, a surviving free function, a per-crateADMITTEDset, an inline match that is not a function). The lanes never disputed the target (one enum,TRAITS[43], inaprender-quant); they dispute that the compatibility shim can be specified completely in advance. C is the same thing seen from the phase table.Found on the orchestrator's re-runs, not by any lane — both stand regardless of the ruling:
GgufTensor::byte_size(crates/aprender-core/src/format/gguf/types.rs:365) sizesQ4_1at 18 bytes/block; upstream is 20. Latent export-size defect, to be filed and fixed on its own.static_assert= 36, gguf-pyGGML_QUANT_SIZES= 40 at llama.cpp3173a56471c1753650cd806694145ffd6dcace67. The ticket's falsifier ("equals gguf-py for every shared id") cannot be met as worded.Decision requested: rule on A/B — e.g. accept a compatibility shim specified by characterization tests rather than by enumeration, or drop source-compatibility from M1 and let the compiler enumerate the sites. Andon stands: 3/3 by 2026-09-18T12:00Z or the ticket is pulled. No code before 3/3.
🤖 Posted by the PMAT-3430 Claude Code session (fable-5-1). The "stop revising, report the history" instruction reached this session relayed by another session, not typed here; the ruling requested above is the operator's.
noahgift commentedon Sep 17, 2026
Milestone triage report — 2026-09-17 (operator-authorized run; milestone moves only)
Capacity C = 29 [V] — median merged PRs per train over the last 3 tags: v0.66.0 29 · v0.67.0 25 · v0.68.0 99 (
gh pr list --state merged --search 'merged:>=2026-09-04', bucketed by tag commit dates 09-04T21:35Z / 09-10T04:55Z / 09-13T06:29Z / 09-17T06:35Z). 0.69.0 cap = min(C,25) = 25. Note the 99 is an outlier train (4 days, slipped 2 days past the 09-15 due date); the median ignores it.Milestones 0.71.0 / 0.72.0 / 0.73.0 created, due 09-24 / 09-27 / 09-30 (the 3-day step of 0.69.0→0.70.0 continued). Departures from
06x-release-schedule.mdon origin/main: v0.68.0 tagged 09-17T06:35Z vs planned 09-15T06:00Z; 0.69.0 milestone still says due 09-18 while the working deadline is 09-20T06:35Z; existing due dates not edited.Open counts per milestone, before → after [V]
Total open: 417 (413 at snapshot + #3452 #3453 #3454 #3455 opened mid-run and folded in; their revert rows: 3452,pr,(none) · 3453,issue,(none) · 3454,issue,0.70.0 unchanged · 3455,pr,0.69.0 unchanged). After-counts re-read from
gh api …/milestonesmatch the plan exactly; zero tickets ahead of a blocker (checked over the edge list below).0.69.0 — 25/25
#[repr(u32)] GgmlType(43 rows: 35 live + 8 RemTensorStorage— a dense consumer cannot recemain— lGgmlTypevs the vendored `ggmtensor_byte_sizeand ONE refusal site readTRAITSqtype: 0+byte_size: 0may not bTier 3 reasons: #3205+#3204 (P0, linked PR, mini CI leg) · #3395+#3401 (P0, linked PR red only on
present; blocks the 24 CRUX-O children) · #3161 (P1, publish-gate, S: add LICENSE-APACHE) · #3170 (P1, root cause stated, S) · #2730 #2765 #3258 (P0, mechanism stated, M). #3452 and #3455 were opened mid-run and attached to Tier 1 (M1 plan PR; the fix PR for #3445), which pushed #3271 and #3005 (EV 0.9) to 0.70.0 to hold the cap.Alfredo's tickets (13 open)
Flag: #3422 #3423 #3424 were moved out of Alfredo's dedicated no-deadline milestone
Inference dispatch & architecture consolidation(now empty, left open) per the run's rule; his #3423 body argues explicitly for keeping them off the cadence. One-line revert is in the CSV.Top 10 EV that did NOT fit 0.69.0
TRAITS(type × bmake build-reportLoad vs capacity, 0.70–0.73
Load 391 vs capacity 4×C = 116 → 3.4× over. 0.70.0 29 · 0.71.0 29 · 0.72.0 29 · 0.73.0 304 = 29 by EV + 275 overflow (flagged in the CSV
whycolumn; no label applied — relabeling was out of scope). At C=29/train the open backlog is ~14 trains deep before any new arrivals; arrival during this 90-minute run was 4.Stale (>90 d no activity, no repro): 0 [V] — oldest
updatedAtis 2026-08-22 (the CLUS-00x set).Duplicate candidates (no action taken)
#3038↔#3090/#3091 · #2730↔#2899 · #2735↔#2887 · #2830↔#2893 · #2816↔#2990 · #2756↔#2888 · #2699↔#2917 · #2648↔#2913 · #2879↔#2898 · #2775↔#2894 · #3226↔#3287↔#3328 (CUDA test binary SIGSEGV at teardown) · #2614↔#2814. Unlinked row/PR pairs: #2924↔PR #3005 · #2906↔PR #3001 · #3133↔PR #3134.
Close candidates (shipped-train residue, not stale by the rule; no action)
#2986 + PR #2998 (PR triage into 0.66) · #2932 (TAG-0.66.0) · #2970 and the nine
carry-0.67rows #2988–#2996 (0.67 shipped) · #3079 (closes with 0.68.1) · #2307 (bot tracker).Other findings
no-close: #Nparse, GitHub parses 'no-close: #N' as 'close: #N' — #3091 was closed twice by PRs whose bodies said no-close (D0 of the pmat 3.41.0 run); #3351 carries the same line for #3347 #3400) — merging it will close CPU: Qwen3.5 / Qwen3.8 hybrid GGUFs are refused entirely — no fallback exists #3091 a third time.autopilot.sh 3450 wait dryrun, untouched) polls release: 0.68.1 #3450 for MERGED and stops atdryrun; it reads milestone 5 only in itsclosestep (needsopen_issues == 0), which this run does not reach. CPU: Qwen3.5 / Qwen3.8 hybrid GGUFs are refused entirely — no fallback exists #3091 moved to 0.70.0; release: 0.68.1 #3450 left in place.ci_queue_steward.shdefaults to the first open milestone by title — that is0.66.0(open, 0 items).EV inputs
impact: priority label [V]; ×2 user-reported crash/refusal with repro — applied to #3038 #3090 #3091 #3075 #3076 [A]; +1 per blocked open ticket, max 3, from the edge list [V where the body/title states it, A otherwise]; ×1.5 release-gate/Qwen2.5 proof line by title keyword [A]. P: 0.9/0.6 from
closingIssuesReferences+ check rollup [V] (only PR #2794 is fully green;present-only red scored 0.6); 0.4 when the body states a root cause/mechanism (regex) [V-regex]; else 0.2. cost S/M/L: title/label heuristic plus 13 hand overrides, all [A]. Ties: gate flag, then already-in-0.69.0, then lower number. A PR and the open issues it closes move as one unit.Blocker edges used: #3430→#3432 #3433 #3441 #3442 #3443 #3077 · #3433→#3434 #3435 · #3443→#3077 · #3045→#2908 · #3015→#2892 · #3147→#3148 #3156 #3158 · #3149→#3155 · #3150→#3168 · #3401→#3371–#3394 · #3418→#3422 #3424 · #3091→#3090 #3038. Quant/dispatch/MoE-loader set held to 0.70.0+ [A]: #2567 #2604 #2605 #2611 #2880 #3332 #3453 + the pre-slotted Phase 2 rows.
full plan CSV (417 rows)
15 remaining items
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: PP-QUANT epic: #4001 moves it from 0.73's enablers to 0.74's core
noahgift commentedon Sep 23, 2026
milestone REVERTED to 0.70.0 (its state before #4024's move). The quorum on #4024 found this issue is named as must-carry by #3999 and #4001, and the table's own rule says such an issue is an operator DECIDE, never placed by the author. It was moved in error. It is now listed on #3998 as a pending decision: 0.73.0/0.74.0. No new placement was made.
docs(roadmap): correct the triage after the quorum -- 3 moves reverte…
docs(roadmap): record the 9 applied corrections, the #3513 withdrawal…
docs(roadmap): quorum round 9 -- #3714 back to 0.69.1 in flight: the …
noahgift commentedon Sep 23, 2026
milestone → 0.74.0, applying the operator's decision on #3999 (2026-09-23 10:43Z): "PP-QUANT (#3421/#3420), PP-ARCH (#3424) and PP-TENSOR (#3428) move to 0.74's epic as its core; 0.73 stays pure performance parity with llama.cpp." Nothing later in the thread reverses it. My earlier revert applied the table's 'two epics name it → DECIDE' rule mechanically, while the operator had already decided. #3999's body still lists these rows and is stale. (#4024 quorum round 11.)
docs(roadmap): quorum round 11 -- a reproducible release-tree census …
noahgift commentedon Sep 25, 2026
Superseded by EPIC E7 #4001 (APR-EPIC-001 A4). Open children re-parented: none. Owner lines copied to the absorbing epic's body. Applied by aprender-d1: claude+agy both said close, apr shadow row recorded (handoff/close-quorum-2026-09-25.json).
PMAT-3430 M1: one ggml tensor-type table, extracted from upstream rat…
PMAT-3430 Phases 3-4: three ggml enums become one (M1 complete) (paim…