docs(roadmap): milestone triage PROPOSAL — 0.70.0 + unmilestoned (nothing moved)

[noahgift]

Summary

A milestone triage of every open issue in 0.70.0 (178) and every open issue with no milestone (136). Each has a milestone and a one-line reason. The operator approved the moves and, later, the corrections (both relayed by the cop session), and they were applied on GitHub with a comment on each issue. The doc records every applied move, correction, and still-pending row. This PR is docs only.

File: docs/roadmaps/milestone-triage-2026-09-23.md

How rows were placed

1. Named must-carry in a theme epic (EPIC 0.70.0: Fast Train — freeze→publish in ≤4 h; lock scoped to GPU work, sharded sweep, idle same-arch pre-screen #3998, EPIC 0.71.0: Don't Leave Behind — every Qwen, every quant, every verb, on CUDA + CPU + wgpu + Apple Silicon, zero REDs #3994, EPIC 0.72.0: Agent Ready — apr serve is the one gateway for all agents (OpenAI + Anthropic /v1/messages), one conformance suite #4000, EPIC 0.73.0: llama.cpp Parity — decode/prefill/TTFT/memory ≥ llama.cpp across the certified matrix #3999, EPIC 0.74.0: Any Model — unified model support (Alfredo #3423/#3422/#3418), new architecture ≈ config + small delta, llama.cpp as reference #4001, EPIC 0.75.0: CRUX declarative fine-tune/distill for Qwen3.5 — one YAML recipe, certified vs 3–5 competitor engines #4002) → that release. An issue named by two epics becomes DECIDE.
2. Otherwise → the release whose exit bar the issue blocks.
3. Debt → a debt-ratchet (EPIC: debt ratchet 0.70→0.74 — 80% of tech debt in 5 equal slices (coverage→95% w/ yoga CUDA shards, pv deepest enforcement, ONT-001 fully merged, zero backlog) #3997) slice by pillar.
4. Non-milestone outcomes: 0.69.1 (in flight), verify-close (the release tree implements it), close? (superseded epic), none (pinned).
5. Every row was hand-checked after a keyword pass; about 120 reasons are hand-written.

Operator decisions surfaced

• IQ2_XXS (ggml 16) has no CUDA GEMV — 0.69.1 release blocker #3950, IQ2_S (ggml 22) has no CUDA GEMV — 0.69.1 release blocker #3953, Q2_K (ggml 10) has no CUDA GEMV — 0.69.1 release blocker #3960, IQ3_XXS (ggml 18) has no CUDA GEMV — 0.69.1 release blocker #3963 (low-bit CUDA GEMV): 0.70 vs 0.71. EPIC 0.70.0: Fast Train — freeze→publish in ≤4 h; lock scoped to GPU work, sharded sweep, idle same-arch pre-screen #3998 and EPIC 0.71.0: Don't Leave Behind — every Qwen, every quant, every verb, on CUDA + CPU + wgpu + Apple Silicon, zero REDs #3994 both name them.
• qwen3moe CUDA reaches run+qa only: chat rc 8, serve 501/500, code rc 1 — serve loads no mapped model and would route MoE to the CPU path #3987 (qwen3moe verbs): 0.70 / 0.71 / 0.72. Three epics name it.
• apr code hardcodes 'apr serve --gpu' (no CPU lane, no --max-tokens/thinking flag) and picks a colliding port 19384+pid%1000 #3978, apr serve: APR-CPU fallback and safetensors routers have no GET / route index and SSE ends without finish_reason/[DONE] — every .apr/.safetensors serve cell is RED #3979 (apr code / serve routes): 0.71 vs 0.72.
• arch constraints: qwen35moe (Qwen3.5-35B-A3B) has is_moe=false; replace #3992's substring MoE predicate with a table entry + case table #4008 (qwen35moe arch constraints): goes with 0.70.0: qwen35moe (Qwen3.5-35B-A3B hybrid SSM MoE) CUDA forward #3977, which is EPIC 0.73.0: llama.cpp Parity — decode/prefill/TTFT/memory ≥ llama.cpp across the certified matrix #3999's own open decision (0.73 vs 0.74).

Notable findings

• Qwen3.5-0.8B-IQ4_XS never closes its think block while Q4_K_M does: CUDA IQ4 GEMV or real quant damage? #3951 (IQ4_XS think block never closes): aprender-83's F9 oracle shows the block closes on the official template, so this is likely resolved by Qwen2.5 chat template drops the default system prompt (26 vs llama.cpp's 47 tokens) — apr 1/18 vs llama.cpp 10/18 on the same GGUF #3990. Re-measure before placing it.
• Six issues are already implemented in the release tree and are proposed verify-close: RELEASE GATE HOLE: check_model_ladder ignores qa_rc / gates_failed — a Q4_K row with apr qa exit 5 is judged ok #3898, ladder_serve_teardown reads 'not answering /health' as 'server gone', so a still-loading server is reported teardown: clean #3943, model_ladder.sh: serve_rc is a dead variable — every serve route can 503/500 and the row is still green #3886, model_ladder receipt: per-row green is not the gate's verdict — 3 rows marked green that check_model_ladder FAILs (mirror of #3842) #3897, P0: chat/code judged by rc and serve by http — a verb emitting 'zombie zombie zombie' is recorded as working, and a GREEN row proves it #3921, CRUX judge floor (#3739): a cell judged on fewer than two engines is UNJUDGED, and the receipt states per cell WHICH engines ran and at WHICH versions #3832.
• Seven superseded or stale rows are proposed close?: Epic: dogfood audit of apr 0.63.0 from crates.io — 201 defects, 24 P0 #2373, EPIC: release train 0.70.0 — receipts replace refusals #3081, EPIC: nvidia-cuda-rust-library-integration.md — stabilize GPU quality with NVIDIA's CUDA Rust libraries (0.67) #3062, OXIDE-001 (0.68.3, fast-track): cuda-oxide pure-Rust #[kernel] → PTX — one GDN kernel ported and measured on lambda-labs, gx10, yoga and mini; every kernel an ont:Kernel with parity, timing and safety receipts per silicon under pv lint --gate shapes #3522, EPIC: technical debt — gates that cannot fail, and the coverage gap #3863, DECISION NEEDED (published for refutation): two 0.69.1 deferral questions — a gate's inspection limit vs a model that does not run #3900, release/0.69.1-batch-1: check_release_bump_pr_body.sh fixture ladder receipts are the pre-#3712 shape — the real judge no longer passes them (combination break of #3721 + #3738) #3770.

keep-open: this PR records milestone placements and changes no code. Every issue it cites stays open under its own milestone, and the verify-close / close? rows wait on their own re-measurement or operator ruling, never on this merge.

Refs #3998

🤖 Generated with Claude Code

[noahgift]

Quorum verdict: AGREED 3/3 PASS, degraded: same-family

• Head judged: 834144822
• Author: claude-opus-5-5
• Lanes: 3 × Sonnet 5 (claude-sonnet-5) via Claude Code, a different model id from the author.
• Why degraded: agy/Gemini quota is exhausted, so the operator's standing rule applies: "if agy quota is ever gone, simply use claude code itself", and same-family quorums "may arm merges, marked degraded: same-family". The first Opus 5.5 lanes were self-review under the operator's model-id rule, so they were discarded and re-run on Sonnet 5.

lane	model	verdict	blocking findings
1: live state, consistency, rulings, census	claude-sonnet-5	PASS	none
2: placement quality and honesty	claude-sonnet-5	PASS	none
3: full independent check	claude-sonnet-5	PASS	none

Non-blocking note from lanes 1 and 3: the post-head drift is large. Other sessions' triage (aprender-dd, from 16:18Z) moved many issues after this snapshot. The doc discloses this, and says the DECIDE rows those passes placed are not rulings.

How the doc got here: earlier rounds (Opus lanes, used as grill passes) found and fixed:

• 5 moves made in error, which were reverted or re-applied on the LATEST operator rulings;
• 10 keyword misplacements, 9 of which were corrected on operator approval (DP4A Q4K/Q6K GEMV profile is catastrophic through the Gated DeltaNet recurrence — Qwen3.5 on sm_89 needs the float MWV variant (PMAT-806 class) #3513 withdrawn);
• 4 duplicates, closed on operator approval;
• a reproducible release-tree census;
• count timestamps that had been written by hand.

Arming with --match-head-commit 834144822.

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-3998",
 "head": "1678c1770b1fe240c68e83a5a78bead99d4ec82d",
 "width": 3,
 "executor": "agy",
 "agreed": false,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "FAIL",
   "findings": 2
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[noahgift]

PMAT-3998 receipt: milestone triage (PR #4024), head after 1678c17

What changed since the last agreed round

One row: #3981's title column no longer quotes its two rate literals, which
check_no_claim_literals.sh bans on a doc surface. The PR body gained a keep-open: line for R-2
(check_pr_closes_issue.sh). Nothing else.

Answers to round-2 lane 1 (gemini-3.1-pro-high)

1. "EPIC: PP-QUANT-001 — quant-type dispatch, Phase 0/1 (0.69 train) #3421/EPIC: PP-TENSOR-001 — a tensor that has no bytes is a different type from one that does (MoE / tied-embedding placeholders) #3428 placed by the author, line 185". Line 185 is The fleet-pinned pv is 0.65.2 and has no 'extract' subcommand and no '--gate' flag — every SHACL gate measured on the fleet binary is unrunnable, not merely unmeasured #3567's row. The EPIC: PP-QUANT-001 — quant-type dispatch, Phase 0/1 (0.69 train) #3421/EPIC: PP-TENSOR-001 — a tensor that has no bytes is a different type from one that does (MoE / tied-embedding placeholders) #3428 rows are
   lines 255-256. Their reason column says the placement is the OPERATOR's decision on EPIC 0.73.0: llama.cpp Parity — decode/prefill/TTFT/memory ≥ llama.cpp across the certified matrix #3999
   (2026-09-23 10:43Z, "PP-QUANT/PP-TENSOR move to 0.74"), found in round 11. The author's own
   placement was the one REVERTED in round 3. AC2 forbids the author placing a two-epic row. It
   does not forbid recording the operator's placement.
2. "AC3: DECIDE/in-flight rows did not stay untouched". Correct as history, and the doc says so
   itself (line 7): EPIC: debt ratchet 0.70→0.74 — 80% of tech debt in 5 equal slices (coverage→95% w/ yoga CUDA shards, pv deepest enforcement, ONT-001 fully merged, zero backlog) #3997/0.70.0: apr code has no backend selector — it spawns its own apr serve, so "code on cpu" vs "code on cuda" is not expressible #3845/qwen3moe (Qwen3-30B-A3B, Qwen3-Coder-30B-A3B, Q4_K_M) has NO working CUDA path: apr run --gpu falls back to CPU (rc 14), apr qa emits no gates — blocks 0.69.1 under the all-Q4_K-CUDA rule #3714 were moved in error and reverted, each with a comment on the
   issue. Every one of them is back in its pre-triage milestone, which is the state AC3 protects;
   the census on the doc shows it. The PR delivers the table and its disclosure, not the error.

[noahgift]

quorum-review (AD-04): NOT agreed (auto_merge: checked=true was_armed=false disarmed=false)

{
 "ticket": "PMAT-3998",
 "head": "1678c1770b1fe240c68e83a5a78bead99d4ec82d",
 "width": 3,
 "executor": "agy",
 "agreed": true,
 "auto_merge": {
  "checked": true,
  "was_armed": false,
  "disarmed": false,
  "note": "auto-merge not armed"
 },
 "lanes": [
  {
   "lane": 1,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 2,
   "verdict": "PASS",
   "findings": 0
  },
  {
   "lane": 3,
   "verdict": "PASS",
   "findings": 0
  }
 ]
}

[noahgift]

Armed by the cop (aprender-cf). Round 3 is 3/3 gemini PASS and the artifact reads agreed:true, lint ok. The script's NOT AGREED / exit 1 comes only from partial:true, caused by agy's idle narration line on lane 1's stderr ("root agent idle; waiting up to 25m0s …"). That is the known tooling defect paiml-implement#352, fixed by #356. It is not a verdict.

[noahgift]

superseded by aprender-dd 0.70 triage (2026-09-24)