aliasfoxkde · Oct 19, 2026 · Oct 5, 2026 · Oct 5, 2026 · Oct 5, 2026 · Oct 5, 2026
diff --git a/.github/CHANGELOG.md b/.github/CHANGELOG.md
@@ -8,6 +8,10 @@ releases may break, and every breaking change is recorded in this file
 
 ## [0.6.1] — 2026-10-15
 
+> Version note: `v0.6.0` was cut as a draft release and its tag pattern was
+> retired (immutable tombstone) before publication; everything it contained
+> shipped in `0.6.1`. There is no published 0.6.0.
+
 The engine learns when not to answer: selective-prediction evidence
 floors and elicited abstention join the calibration stack, the corpus
 toolchain becomes a first-class item generator crate, and the

diff --git a/.github/SECURITY.md b/.github/SECURITY.md
@@ -4,7 +4,9 @@
 
 | Version | Supported |
 | ------- | --------- |
-| 0.1.x   | yes       |
+| 0.6.x   | yes       |
+| 0.5.x   | yes       |
+| < 0.5   | no        |
 
 ## Reporting a vulnerability
 

diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
@@ -13,6 +13,7 @@ in a committed phase, **contract** = frozen v1 surface.
 | Schema adapters | `opencodifier-schema` | core |
 | Inference runtime | `opencodifier-runtime` | core (+ `ort` behind one feature) |
 | Decision model | `opencodifier-model` | engine, runtime, core |
+| Item generator | `opencodifier-itemgen` | core, engine |
 | Interfaces | `opencodifier-cli`, `-http`, `-mcp`, `-wasm`, `-ffi` | engine/model, schema |
 
 The dependency arrows only ever point toward `core`. Nothing in the

diff --git a/docs/DECISIONS.md b/docs/DECISIONS.md
@@ -148,7 +148,7 @@ re-ordering, candidate narrowing subset-safety. Fixtures under
 | `runtime-onnx-dynamic` | opencodifier-runtime | ort `load-dynamic` |
 | `llamacpp` | opencodifier-model | ureq (HTTP client only, D26) |
 | `tokenizers` | opencodifier-model | `tokenizers` 0.22 + `fancy-regex` (Kai reference encoder, D33) |
-| `http` | opencodifier-http | axum 0.9, tokio |
+| `http` | opencodifier-http | axum 0.8, tokio (pin corrected by D12) |
 | `mcp` | opencodifier-mcp | rmcp 2.2 (D1) |
 | `cli` | opencodifier-cli | clap v4 |
 | `wasm` | opencodifier-wasm | wasm-bindgen, getrandom js |

diff --git a/docs/PLAN.md b/docs/PLAN.md
@@ -62,7 +62,9 @@ Remaining risks tracked in §6 below.
 | 19 | Adoption + integration documentation | **done** (2026-10-04) — `docs/INTEGRATIONS.md` + D30/D31 + 19e Amortyx E2E |
 | 20 | Coverage restoration: the escalation core | **done** (2026-10-08) — 20a ratchet done (D36); 2026-10-07 census re-based on the pinned image (**98.03 % in-image DA**, 350-line residual; the 980-line "executor debt" was a local-mapping artifact), floor 89.0 → 97.5, COVERAGE.md regenerated, 20h done; 20g tranches 1–3 done 2026-10-08 (cli ladder shell → itemgen probe/guard arms + UTF-8 boundary → macjev fall-through): **98.39 %** (18 187 / 18 485), residual 298, every line in a named infeasible class — unit-reachable pool exhausted, D36 ratchet retired at the reachable ceiling, floor rests at 97.5 fixed |
 | 21 | Site WCAG 2.1 AAA + progressive enhancement | **opened** (2026-10-06) — audit + fixes landed; deploy + X/Discord pending |
-| 22 | Docs freshness sweep | **opened** (2026-10-06) |
+| 22 | Docs freshness sweep | **in progress** (2026-10-06 opened; 2026-10-15 sweep: SECURITY matrix, D11 axum pin, §36 route note, itemgen in ARCHITECTURE, site counts, wiki cross-checks) |
+
+Handoff of record: `docs/handoffs/PROGRAM_HANDOFF_2026-10-15.md` (state, next steps, ranked improvements, roadmap incl. the vision rung; supersedes no record, organizes them).
 | 23 | v0.5.0 release cut | **done** (2026-10-08) — tag `v0.5.0` → `0fd6a154`, 9-leg matrix rebuilt single-tree with one consistent attestation set; releases on GitForge + GitHub (19 assets, 69.4 MB) |
 | 24 | RESEARCH §15.6 enhancement program | **in progress** (2026-10-07) — items 1–4, 8–10 + §15.8 fold-ins done (ledger: RESEARCH §15.6.1); items 5/6 sized, model-rung options, deferred per the §15.6 note |
 

diff --git a/docs/PLANNING.md b/docs/PLANNING.md
@@ -1488,6 +1488,13 @@ POST /v1/systemone
 
 The compatibility endpoint must emulate the supported Jev/System One schema, while the native `/v1/decide` API remains the canonical OpenCodifier API.
 
+*Amendment (2026-10-15): shipped reality — there is no `/v1/systemone`
+route. Jev/System One compatibility lives in `opencodifier-schema` as a
+wire adapter (Jev in, canonical IR out), and the HTTP surface is
+`POST /v1/decide` (+ `/v1/chat/completions`, `/v1/messages`,
+`/v1/batch`). The adapter-normalizes-into-the-IR rule supersedes a
+dedicated compat endpoint; see docs/DECISIONS.md D12.*
+
 ---
 
 # 37. Local Server

diff --git a/docs/PROJECT_STRUCTURE.md b/docs/PROJECT_STRUCTURE.md
@@ -73,12 +73,15 @@ opencodifier/
 │   ├── *-core.js            # one DOM-free, node-testable core per game
 │   │                        #   (pong, life, pacman, racing, fighter, quiz,
 │   │                        #   tower, antcolony, civ60, traffic, spacecraft,
-│   │                        #   dilemma, rpsls, door) — seeded, deterministic
+│   │                        #   dilemma, rpsls, door, cube + cube-n, clock,
+│   │                        #   pyraminx) — seeded, deterministic; the twisty
+│   │                        #   cores carry proved solvers, not heuristics
 │   ├── <game>.js            # per-game rendering shells (canvas/DOM chrome;
 │   │                        #   rendering never touches an rng stream)
 │   ├── playground.html      # merged playground: maze demo + game library
 │   └── tests/               # node --test suites: board page, maze
-│                            #   core/shell/deep-link, SDK games (20 files)
+│                            #   core/shell/deep-link, SDK games, cube cores,
+│                            #   clock (27 files)
 └── docs/                    # PLANNING.md (founding), PLAN.md (live plan), …
                              #   RELEASE_NOTES_TEMPLATE.md is the release-notes
                              #   skeleton; `just release-notes` fills its

diff --git a/docs/handoffs/PROGRAM_HANDOFF_2026-10-15.md b/docs/handoffs/PROGRAM_HANDOFF_2026-10-15.md
@@ -0,0 +1,173 @@
+# Program Handoff — 2026-10-15
+
+**Status:** handoff document. It organizes the state of record and names the
+next steps with their stopping conditions. Every factual claim here is a
+pointer to a record of record (listed per item); nothing in this document is
+itself primary evidence. Times are US Central.
+
+**Branch of note:** `site/games-audit-2026-10` (pushed to GitForge and
+origin) carries the 2026-10-15 site work: cube overhaul (`ce9d65a`) and the
+docs freshness sweep (`553cd28`). Site flow: edits on branch → Cloudflare
+Pages preview → human merges the PR.
+
+---
+
+## 1. State of record
+
+**Release.** `v0.6.1` (2026-10-15): selective prediction end to end,
+`opencodifier-itemgen` as a crate, corpus-composition question closed, first
+gated weights on HuggingFace. `v0.6.0` is an unpublished tombstone (tag
+pattern retired); `.github/CHANGELOG.md` says so explicitly. Coverage 98.39 %
+in-image DA; CI of record is GitForge on the fedora remote.
+
+**Model program** (`docs/TRAINING.md`, all verdicts recorded there):
+
+| Arm | Verdict | Numbers of record |
+|---|---|---|
+| E1-E (2B, iterb scaffold 178k) | **GATE_PASS ×2 — release candidate** | suite 0.8083, JevBench 0.6840, ECE 0.127 (T=3.362) |
+| t33p (2B, merged-v6-full 307k) | GATE_PASS ×2 — evidence, not candidate | suite 0.6917, board 0.6580, lane drop −3.4 (vs E1-E's −12.4) |
+| T88 imatrix Q4_K_M quant | **FAIL all three gates** | recorded 2026-10-15, 10:20 AM |
+| MiniCPM5 GGUF rung | CLOSED no-go | adapter lane-bound; size/speed legs win |
+| Corpus-volume lever | **CLOSED** | v5 catastrophic → v6 gate-passing → scaffold alone best; "gated volume ≠ transfer" |
+
+The two legs of the release gate: suite ≥ 0.68 AND JevBench ≥ 0.6494
+(`docs/RELEASE_POLICY.md`; [[Models]] on the wiki mirrors it). Weights
+publish only gate-gated to HuggingFace (`TaskWizerAI`), open Apache-2.0.
+
+**Site / playground** (`site/`, branch `site/games-audit-2026-10`):
+
+- **Twisty Cubes 2×2–21×21** (`cube.html`, `cube.js`, `cube-core.js`,
+  `cube-n-core.js`): two verified cores under one facade. 3×3 keeps the
+  battle-tested LBL core; 2×2 has its own coach (layer build → (R′D′RD)
+  orientation → BFS over the 24 top permutations); 4×4–21×21 are validated
+  free play with piece-law checks on every move and SiGN wide turns.
+  Renderers: CSS-3D orbit at n ≤ 3, canvas net at n ≥ 4. Mirror/ghost/
+  sudokube skins at 3×3. The long-standing T-perm typo (15 tokens) is fixed,
+  and the greedy phase-3 stall is replaced by exhaustive BFS. Tests:
+  cube-n-core 13, cube-core 10, playwright smoke 25/25.
+- **Rubik's Clock** (`clock.html`, `clock.js`, `clock-core.js`): TNoodle-pinned
+  move table, rank-14 basis over Z₁₂, 12¹⁴ reachable states, **unique
+  solution** — solved by exact elimination, replay-verified; engine-free on
+  purpose (no ties exist to arbitrate). Tests: core 12/12, smoke 13/13.
+  Two shell bugs found by smoke and fixed: plan moves no longer invalidate
+  their own plan mid-step, and the boot chip now renders (and truthfully
+  reads "scrambled" — the page boots paused per the site convention).
+- **Playground hub** (`playground.html`): truthful card badges — playable vs
+  "verified core is in test" (Pyraminx, Skewb) vs "designed, not faked"
+  (Megaminx, Square-1, Fisher, Windmill).
+- **Docs sweep** (`553cd28`): SECURITY supported-matrix → 0.6.x; D11 axum pin
+  corrected to 0.8; PLANNING §36 amendment (no `/v1/systemone` route — Jev
+  compat ships as a schema adapter); itemgen added to ARCHITECTURE; site
+  counts corrected (27 test suites); PLAN phase 22 notes the sweep;
+  CHANGELOG explains the 0.6.0 tombstone.
+- **Wiki** (GitHub wiki, pushed `ae89b4e`): 17 pages + footer/sidebar;
+  Roadmap now carries the vision rung (§3 below); Playground lists the cube
+  family and Clock truthfully.
+
+**Compute posture** (memory + `docs/handoffs/PLATFORM_DECISION_RUNTIME_ADOPTION_HANDOFF_2026-10-12.md`):
+the T5500 is down; fedora carries the eval/training load through harness-jobs;
+local hosts serve models; training is free-tier-first (Kaggle T4) with a ~$50
+Modal burst budget authorized. CI runs on GitForge, never local cargo.
+
+## 2. Immediate next steps (in order)
+
+1. **Integrate Pyraminx + Skewb** — cores are being built by the in-flight
+   agent (site/pyraminx-core.js exists; Skewb pending). Stopping condition:
+   both cores pass node suites (TNoodle/`jaap`-pinned move tables, seeded
+   solve sweeps), pages ship (`pyraminx.html`, `skewb.html`), playground
+   "soon" badges flip to playable. Do NOT fake a solver: if the generator
+   search can't prove a basis, the page ships free-play with laws validation,
+   like the big cubes.
+2. **Games-audit PR through review** — push any remaining branch work, open
+   the CF Pages preview, and hand the operator the PR. Stopping condition:
+   human merge to main + production re-deploy.
+3. **Board submission** — `docs/board-submission-draft.md` + 8 operator
+   questions are the blockers; the dominance verdict (composite plausible,
+   every-axis not yet) is recorded in `docs/BENCHMARKS.md`. Stopping
+   condition: operator answers → submission filed.
+4. **#150 eval lane v2** — batching + sdpa with anchor-drift check. t33p's
+   −3.4 lane agreement (vs E1-E's −12.4) makes the lane-honesty question
+   sharper: the suite leg may be partially lane-specific. Stopping condition:
+   load-pinned determinism replay or an agreed tolerance, recorded in
+   `docs/TRAINING.md`.
+
+## 3. Areas of improvement (ranked)
+
+1. **Vision support — the operator requirement.** "It can be optional and a
+   separate rung/lever that routes only when needed, but we need vision
+   support to compete… And it needs to be comprehensive and fast." Design
+   shape (recorded on the wiki Roadmap, mirrors the binding rules):
+   - `ImageBackend` seam beside `InferenceBackend`/`EmbeddingBackend` in
+     `opencodifier-runtime`; ONNX behind the feature gate (D11); no Python
+     at runtime (§73).
+   - Rung position: above lexical/embedding, below the text decision model;
+     fires **only on image-bearing requests** — the default text path never
+     pays for it.
+   - Outputs are candidate-conditioned typed decisions (choice/boolean/
+     score), never free-form generation; confidence passes the mandatory
+     calibration gate; cache keys carry vision-model + calibration versions.
+   - Models: small quantized vision encoders (SigLIP-class) on the ONNX CPU
+     path, explicit opt-in pull (release policy), verifier contract mirrored
+     from the text ladder (agree → accept, disagree → abstain).
+   - Open work: sizing + first parity fixtures (fixture-first, per the
+     model-import parity discipline in quality rules).
+2. **Big-cube solver honesty.** 4×4–21×21 are free play by design. A
+   reduction solver is a real project; if wanted, size it as its own phase
+   (parity handling is the hard part) rather than diluting the coach.
+3. **Megaminx / Square-1 / Fisher / Windmill.** Designed-but-not-faked on
+   the playground. Megaminx needs a dodecahedral core (12-color laws); 
+   Square-1 needs a non-group-regular move algebra (shape-dependent). Both
+   deserve their own verified cores before any page ships.
+4. **Per-rung calibration artifacts** (D25/F24 seam exists). E1-E's per-kind
+   T fit is the working pattern; ship rung artifacts only where the ECE gate
+   passes.
+5. **Docs phase 22 residue.** The wiki cross-checks ride the same phase;
+   re-run the stale-doc probe after each release cut, not ad hoc.
+
+## 4. Roadmap
+
+- **Near (this cycle):** Pyraminx + Skewb pages; games-audit PR merged; board
+  submission filed; #150 lane v2 spec'd; vision rung sized with parity
+  fixtures named.
+- **Mid:** vision rung implementation behind its feature gate; consensus-flip
+  patch / strands consensus pass (quality-first trainer lever); T3.4 4B
+  contingency if the 2B plateau holds; per-rung calibration artifacts where
+  gates pass; Megaminx core.
+- **Far:** big-cube reduction solver; hosted tier per
+  `docs/HOSTED_TIER.md`; registry publication (explicit operator go, D24);
+  Square-1 core.
+
+**Deliberately not scheduled** (unchanged): default-on escalation ladder
+(per-domain calibration precedes it), browser model path (D23), any
+cloud/API-key dependency (the default posture is fully local and offline).
+
+## 5. Standing constraints (unchanged, binding)
+
+- Push order: `git push gitforge` first, then `origin`; pushed history is
+  immutable — fix forward, never force-push; never `git stash`.
+- CI/CD through GitForge on fedora; a red GitHub Actions run is not a code
+  signal. Heavyweight checks via harness-jobs profiles.
+- Never echo credentials (gitforge/Kaggle/HF/env files); `gitforge auth
+  --login` is the operator's credential path.
+- Conventional commits, `Co-Authored-By: Claude` trailer; aegis baseline
+  regen rides the same commit as source edits.
+- Site conventions: boot-paused games with truthful play buttons; accordion
+  exclusivity site-wide; stats overlay on every game; responsive rule for
+  every canvas; edits on branch → CF preview → human merge.
+- Findings land in repo docs + commit + push before the turn ends; memory
+  holds pointers only.
+- Reports in AM/PM 12-hour US Central.
+
+## 6. Where everything lives
+
+| Record | Path |
+|---|---|
+| Live plan + phase table | `docs/PLAN.md` |
+| Binding decisions | `docs/DECISIONS.md` (D1–D37+) |
+| Founding spec | `docs/PLANNING.md` (§§1–80) |
+| Training record + verdicts | `docs/TRAINING.md` |
+| Benchmark page of record | `docs/BENCHMARKS.md` |
+| Release policy + notes flow | `docs/RELEASE_POLICY.md`, `.github/CHANGELOG.md` |
+| Site release flow | `docs/SITE_RELEASE_FLOW.md` |
+| Wiki | GitHub wiki (17 pages; clone at `/var/tmp/wiki-oc`) |
+| Prior handoff (runtime adoption) | `docs/handoffs/PLATFORM_DECISION_RUNTIME_ADOPTION_HANDOFF_2026-10-12.md` |