feat(skills): implement federated scientific MCP platform

.github/dependabot.yml

@@ -14,11 +14,8 @@
 #     (a Virtual-Scientists fork) and ari-skill-paper-re/vendor/paperbench
 #     (openai/preparedness) are pinned external forks; auto-bumping their
 #     submodule SHAs is undesirable.
-#   * pip "/ari-skill-orchestrator" is NOT listed. That one skill ships no
-#     pyproject.toml (13 of 14 skills have one); pointing Dependabot at a
-#     missing manifest would error. Only first-party directories are listed,
-#     so the vendored pyproject.toml files under the two vendor/ submodule
-#     trees above are also excluded.
+#   * Only first-party directories are listed, so vendored pyproject.toml files
+#     under the two vendor/submodule trees above are excluded.
 #   * docker ecosystem is NOT added. There are no in-tree Dockerfiles outside
 #     vendor/ (containers/ holds only a README).
 #
@@ -45,10 +42,9 @@ updates:
       - "github-actions"
 
   # --- Python (pip) ---------------------------------------------------------
-  # One block covering the 15 verified first-party manifest directories:
-  # the root requirements.txt, ari-core, and the 13 skills that ship a
-  # pyproject.toml. ari-skill-orchestrator is intentionally absent (no
-  # manifest). Minor/patch bumps are grouped per directory to cap PR volume
+  # One block covering the 17 verified first-party manifest directories:
+  # the root requirements.txt, ari-core, and all 15 skills. Minor/patch bumps
+  # are grouped per directory to cap PR volume
   # across the tree; a low open-PR limit adds a second guardrail.
   - package-ecosystem: "pip"
     directories:
@@ -60,10 +56,12 @@ updates:
       - "/ari-skill-hpc"
       - "/ari-skill-idea"
       - "/ari-skill-memory"
+      - "/ari-skill-orchestrator"
       - "/ari-skill-paper"
       - "/ari-skill-paper-re"
       - "/ari-skill-plot"
       - "/ari-skill-replicate"
+      - "/ari-skill-tool-registry"
       - "/ari-skill-transform"
       - "/ari-skill-vlm"
       - "/ari-skill-web"

.github/workflows/contracts.yml

@@ -106,6 +106,17 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: "3.13"
+      - name: Install manifest validation dependencies
+        run: python -m pip install pydantic pyyaml
+      - name: Canonical Skill manifest conformance
+        # Hard gate: an invalid/unversioned manifest, runtime tool drift,
+        # workflow drift, version drift, generated mcp.json drift, or a
+        # default-enabled collision is an admission failure.
+        run: python scripts/check_skill_manifests.py
+      - name: Generated Skill and execution schema drift
+        run: python scripts/sync_skill_metadata.py
+      - name: Generated HPC contract schema drift
+        run: python ari-skill-hpc/scripts/sync_contracts.py
       - name: MCP tool-schema snapshot verify (advisory at Stage 1)
         continue-on-error: true
         run: python scripts/snapshot_contracts.py --surface mcp --check

.gitignore

@@ -30,6 +30,9 @@ slurm-*.out
 # Checkpoint directories (generated at runtime)
 *_ckpt_*/
 checkpoints/
+# Versioned migration fixtures exercise legacy checkpoint readers in CI.
+!/ari-core/tests/fixtures/checkpoints/
+!/ari-core/tests/fixtures/checkpoints/**
 output*/
 results*/
 
@@ -92,6 +95,8 @@ ari-core/checkpoints/
 # ── Runtime / temp files (generated at runtime, never track) ────
 experiment.md
 ari-core/experiment.md
+# The migration contract intentionally contains historical experiment.md.
+!/ari-core/tests/fixtures/checkpoints/v0_7_golden/experiment.md
 tmp*.md
 GUI_TASK.md
 *_TASK.md

CHANGELOG.md

@@ -2,6 +2,35 @@
 
 All notable changes to ARI are documented here. Versions follow `MAJOR.MINOR.PATCH`.
 
+## Unreleased — Skill-platform P6 removal
+
+- Removed the deprecated web provider aliases, mutable backend selector,
+  iterative LLM collector, and `papers`/`results` retrieval projections. Use
+  `search_papers(provider=...)`, canonical `records`, and explicit broker/workflow
+  composition.
+- Removed coding-skill's flat measurement writer and permissive JSON coercion.
+  `emit_results` now writes only the typed `measurement_set`; the old reader is
+  read-only migration support.
+- Removed replicate-skill's low-coverage single-call generator, prompt, API/env
+  switches, GUI controls, and report field. Rubric generation is always
+  `hierarchical-v2`/`calibrated`, with model-call budgets and repair provenance.
+- Removed the five container-specific HPC public aliases and their compilers.
+  Container work uses `container_submit`. The core-agent `slurm_submit` bridge
+  remains narrowly supported until that agent emits `JobRequestV1` directly.
+- Added the permanent
+  [compatibility support policy](docs/reference/compatibility_support.md) with
+  owners and objective re-evaluation gates for retained readers, deployment
+  paths, PaperBench adaptations, and registry repair.
+- Externalized the remaining paper claim-declaration guidance without changing
+  the bytes sent to the model; prompt and composition digests are regression
+  tested.
+- Removed the completed temporary skill-platform plan set after its
+  architecture, contracts, migration rules, and retained-support decisions were
+  transferred to permanent documentation.
+
+Pre-removal rollback boundary: `c487ea9` (`feat(orchestrator): add durable
+authenticated run control`).
+
 ## v0.9.1 — Contract-preserving refactoring program (73 subtasks) + DONE-verification audit (2026-07-05)
 
 - **73-subtask refactoring program, every contract preserved.** ari-core, the 14

README.ja.md

@@ -95,12 +95,12 @@ MCP ツール・描画出力のいずれも変更なし。
   でドッグフードに使用可能。
 - **`container_image` のエンドツーエンド配線** — ウィザード → API ワーカー →
   MCP ツール → サンドボックスランナーまで同じ 1 フィールドが流れる。
-  `pb-env` / `pb-reproducer` の短縮エイリアスは
-  `scripts/build_pb_images.sh` でビルドされる `image:latest` タグに解決。
+  v1.0 は immutable なローカル SIF、完全な Docker
+  `sha256:<image-id>`、または digest 固定 URI のみを許可し、mutable な
+  `pb-env` / `pb-reproducer` エイリアスは削除した。
 - **fail-loud な事前条件チェック** — サンドボックス / GPU の不整合は
-  既定で `RuntimeError` を送出（従来サイレントに CPU 実行へ降格していた
-  4 箇所を修正）。互換挙動は `ARI_PHASE1_ALLOW_FALLBACK=1` と
-  `ARI_SLURM_ALLOW_NO_GRES=1` でオプトイン可能。
+  エラーで停止する。旧 host-local 再現 fallback は v1.0 で削除され、GPU
+  要求も黙って降格できない。
 - **PaperBench env-truth ガードレール** — Stage 1 のプロンプトに
   「scaffold 前にホストを probe する」「言語選択を Python 偏重から
   打ち消す」「ホスト実機を反映した `ADDITIONAL NOTES`（バイナリ / GPU /

README.md

@@ -95,12 +95,12 @@ See [CHANGELOG.md](CHANGELOG.md#v081--structural-refactor-frontend-decomposition
   surface. Drives the dogfood CLI via
   `scripts/sc_paper_dogfood.py --with-rollout / --with-reproduction`.
 - **`container_image` end-to-end** — one field flows wizard → API worker →
-  MCP tool → sandbox runner, with `pb-env` / `pb-reproducer` aliases
-  resolved by `scripts/build_pb_images.sh`.
-- **Fail-loud preconditions** — sandbox / GPU mismatches now raise
-  actionable `RuntimeError`s by default (four silent-downgrade sites
-  fixed); legacy fallbacks behind `ARI_PHASE1_ALLOW_FALLBACK=1` and
-  `ARI_SLURM_ALLOW_NO_GRES=1`.
+  MCP tool → sandbox runner. v1.0 admits only a local immutable SIF, a full
+  Docker `sha256:<image-id>`, or a registry URI pinned by digest; mutable
+  `pb-env` / `pb-reproducer` aliases were removed.
+- **Fail-loud preconditions** — sandbox / GPU mismatches raise actionable
+  errors. The former host-local reproduction fallback was removed in v1.0;
+  GPU requests likewise cannot be silently downgraded.
 - **PaperBench env-truth** — Stage 1 prompts now probe-before-scaffold,
   counter-prime the language choice, and inject a host-truthful
   `ADDITIONAL NOTES` block (binaries / GPU / network / Phase-2 isolation).
@@ -352,7 +352,9 @@ After a run completes, outputs are saved in `./checkpoints/<run_id>/`:
 
 ### Skills (MCP plugin servers)
 
-13 skills total. 12 are registered by default in `workflow.yaml`; 1 additional skill (orchestrator) can be enabled by adding it to the config.
+15 skills total. 13 are registered by default in `workflow.yaml`; 2 additional
+skills (orchestrator and the federated tool registry) are default-off and can be
+enabled explicitly.
 
 In v0.6.0 two skills were retired: `ari-skill-figure-router` was folded into `ari-skill-plot` (a single skill now owns both matplotlib plots and SVG architecture diagrams, both feeding the same VLM review loop), and `ari-skill-review` (rebuttal generation) was deleted — the rubric-driven review score is the final quality signal.
 
@@ -370,7 +372,9 @@ In v0.6.0 two skills were retired: `ari-skill-figure-router` was folded into `ar
 | `ari-skill-benchmark` | CSV/JSON analysis, plotting, statistical tests | ✗ | ✓ |
 | `ari-skill-vlm` | Vision-Language model figure/table review | ✓ | ✓ |
 | `ari-skill-coding` | Code generation + execution + file read + bash | ✗ | ✓ |
+| `ari-skill-replicate` | PaperBench-compatible rubric generation and audit | ✓ | ✓ |
 | `ari-skill-orchestrator` | Expose ARI as MCP server, recursive sub-experiments, dual stdio+HTTP transport | ✗ | — |
+| `ari-skill-tool-registry` | Immutable federation, scientific admission, and replay for large MCP collections | ✗ | — |
 
 ✗ = no LLM, △ = LLM used in some tools only, ✓ = primary tools use LLM.
 
@@ -425,4 +429,3 @@ What makes this paper different is not a headline number but a property: **every
 ## License
 
 MIT. See [LICENSE](LICENSE).
-

README.zh.md

@@ -89,12 +89,11 @@ ARI 围绕一个原则设计：**用 Markdown 描述目标 — 其余的交给 A
   `scripts/sc_paper_dogfood.py --with-rollout / --with-reproduction`
   在自测中驱动。
 - **`container_image` 端到端贯通** — 同一个字段从向导 → API worker →
-  MCP 工具 → 沙盒运行器一路传递；`pb-env` / `pb-reproducer` 短别名
-  通过 `scripts/build_pb_images.sh` 解析为对应的 `image:latest` 标签。
-- **失败时立即报错的前置检查** — 沙盒 / GPU 不匹配的四处原本静默
-  降级到主机 CPU 的位置，现在默认抛出可操作的 `RuntimeError`；
-  兼容回退路径通过 `ARI_PHASE1_ALLOW_FALLBACK=1` 与
-  `ARI_SLURM_ALLOW_NO_GRES=1` 显式启用。
+  MCP 工具 → 沙盒运行器一路传递。v1.0 仅接受不可变的本地 SIF、完整
+  Docker `sha256:<image-id>` 或按摘要固定的 registry URI；可变的
+  `pb-env` / `pb-reproducer` 别名已删除。
+- **失败时立即报错的前置检查** — 沙盒 / GPU 不匹配会直接报错。旧的
+  host-local 再现回退已在 v1.0 删除，GPU 请求也不能静默降级。
 - **PaperBench env-truth 护栏** — Stage 1 提示加入
   「先 probe 再 scaffold」「对抗语言选择的 Python 偏置」
   「按主机实测注入 `ADDITIONAL NOTES`（二进制 / GPU / 网络 / 第二阶段隔离）」