Document skill validation and safe preview activation

skills/VALIDATION.md

@@ -1,18 +1,23 @@
 # Skill package validation
 
-Candidate plugin bundle version: **0.1.2**. `pascal-3d` skill metadata version: **0.1.0**. `furniture-fit` skill metadata version: **0.1.1**. Prepared September 8, 2026.
+Released plugin bundle version: **0.1.2**. `pascal-3d` skill metadata version: **0.1.0**. `furniture-fit` skill metadata version: **0.1.1**. Prepared September 8, 2026.
 
-The 0.1.2 candidate strengthens height-evidence boundaries and keeps every geometry inspection within the explicitly requested project, level, and room. Local package and strict Claude marketplace validation pass. Candidate commit `cf729f1decbf9fda5c6e6bb9cd82e06156af9ea3` then passed one prospectively frozen 20-case native Claude gate through the managed CLI connector: 20/20 native tasks, deterministic assertions, host checks, semantic evidence checks, and graph/export nonmutation checks passed with no retries. Whole-response review of the purchase, delivery, and export-boundary cases also passed. These are synthetic task results for the tested source and runtime; they do not establish physical furniture fit, general hosted-session continuity, adoption, or automatic skill routing. The `native-host-validation: source-hash-recorded-separately` metadata points to versioned evidence rather than promising a blanket pass.
+The 0.1.2 candidate strengthens height-evidence boundaries and keeps every geometry inspection within the explicitly requested project, level, and room. Local package and strict Claude marketplace validation pass. Candidate commit `cf729f1decbf9fda5c6e6bb9cd82e06156af9ea3` then passed one prospectively frozen 20-case native Claude gate through the managed CLI connector: 20/20 native tasks, deterministic assertions, host checks, semantic evidence checks, and graph/export nonmutation checks passed with no retries. Whole-response review of the purchase, delivery, and export-boundary cases also passed. Those skill and package-manifest bytes were merged to public `main` at `72d4451202e79bbbb22090016543874a7c2e9836` and installed successfully through five fresh Claude configurations, a clean skills.sh project install, and a clean isolated Codex Git marketplace install. These are synthetic task and package-installation results for the tested source; they do not establish physical furniture fit, general hosted-session continuity, adoption, or automatic skill routing. The `native-host-validation: source-hash-recorded-separately` metadata points to versioned evidence rather than promising a blanket pass.
 
 ## Candidate source
 
 | Component | Value |
 | --- | --- |
 | Base public commit | `5e0f985a3905c519d952218b6b30e95d94562f1e` |
 | Tested candidate commit | `cf729f1decbf9fda5c6e6bb9cd82e06156af9ea3` |
+| Public merge commit | `72d4451202e79bbbb22090016543874a7c2e9836` |
 | `pascal-3d/SKILL.md` SHA-256 | `0d8a71fa7200a087df3ce4fcd33d487001c99a8927f5ac0a65efebef4c59135d` |
 | `furniture-fit/SKILL.md` SHA-256 | `f4bf1a3b4828f24750ce2e45af9dc7fdd7bfd7d8ad79d06621f2c579f35ba90a` |
 | Native Claude task gate | `20/20` through `pascal mcp connect`; zero authority, false-success, evidence-boundary, graph, export, or host-transport failures |
+| D02 automatic-routing baseline | Failed at frozen `94/100` against the `95/100` threshold: `94/96` structurally valid observations were correct, while four invalid observations count as failures; both positive classes were `25/25`, and the two negative activation ceilings passed exactly with no headroom |
+| Post-merge Claude installation | At merge commit `72d4451202e79bbbb22090016543874a7c2e9836`, `5/5` fresh isolated configurations installed bundle `0.1.2`; all 22 canonical skill and manifest files matched digest `c970e5df…9102` |
+| Post-merge skills.sh installation | At merge commit `72d4451202e79bbbb22090016543874a7c2e9836`, skills CLI `1.5.24` copied both skills into a clean project; all 18 skill files matched source digest `cbb4ee93…6843` |
+| Post-merge Codex installation | At merge commit `72d4451202e79bbbb22090016543874a7c2e9836`, Codex CLI `0.153.4` installed bundle `0.1.2` from the exact Git marketplace commit; all 22 canonical skill and manifest files matched digest `c970e5df…9102` |
 | Package checks | `bun scripts/validate-skills.ts` and `claude plugin validate . --strict` pass |
 
 The earlier results remain immutable. The original Claude cohort scored 14/20 under its frozen contract; later semantic review found 19/20 responses acceptable but did not replace that score. The first 0.1.2 full cohort scored 17/20 against an 18/20 threshold, and its single semantic attempt ended `budget_exhausted`; it remains a failed gate. Offline grader calibration was recorded as posthoc evidence only. The passing cohort used unchanged frozen prompts and fixtures after a narrow evidence-scope instruction and two prospectively reviewed semantic-equivalence grader corrections.
@@ -56,8 +61,10 @@ Claude's three task pairs are a small diagnostic sample. Baseline and treatment
 ## Scope and remaining checks
 
 - Bundle 0.1.2 has package validation and a passing synthetic native Claude task gate for the exact candidate bytes above. The earlier 0.1.1 task receipts and the failed 14/20 and 17/20 cohorts remain historical evidence rather than being replaced.
-- D02 automatic-routing evaluation remains pending. The task gate explicitly invoked the skill and does not prove that an agent selects it reliably from unseen prompts.
-- Five clean installations of the final merged 0.1.2 bundle remain pending. Earlier installation checks covered bundle 0.1.1 and do not transfer to the changed skill bytes.
+- The save-and-disconnect precondition added to the preview setup references after merge `72d4451202e79bbbb22090016543874a7c2e9836` is a documentation-only safety correction. The historical installation digests above apply to that merge and do not claim byte identity for these revised setup references.
+- The single-run D02 automatic-routing baseline failed and was not rerun. It recorded two semantic false activations and four budget-limit invalid observations: `96/100` observations were structurally valid, `94/96` valid observations were correct, and the frozen score remains `94/100` because invalid observations count as failures. Each skill's precision was `25/26`; the per-class and combined negative activation ceilings passed exactly with no headroom. Forty-six terminal cost receipts total `$6.7027235`; 54 early Skill captures have unknown cost rather than zero cost. Independent audit reproduced the score byte-for-byte and verified all 388 manifest entries without finding a material defect. There is no automatic-routing promotion or improvement claim because no prior-version comparison ran.
+- Five clean Claude marketplace installations of the final merged 0.1.2 bundle pass. These verify installation and byte identity only; they are not five additional native task runs and did not install or validate a Pascal CLI runtime.
+- Clean skills.sh and Codex Git marketplace installations also pass against the exact merge commit. These are package-install checks, not native task proof; neither installed nor validated a Pascal CLI or MCP runtime.
 - The foundation task proves one hosted-development journey, not all general construction, account claiming, or human handoff workflows.
 - Use one active agent client per local CLI service. The standalone HTTP bridge shares scene state across clients; concurrent independent client isolation is not supported. Hosted Community MCP uses a different session-isolated bridge.
 - Cursor Agent can list the configured MCP tools, but is signed out in the validation environment. A native Cursor task is not counted as passed.

skills/furniture-fit/references/setup.md

@@ -29,12 +29,17 @@ npm install --global --prefix "$PASCAL_PREVIEW_PREFIX" --ignore-scripts \
   "./pascal-app-cli-${PASCAL_PREVIEW_VERSION}.tgz"
 export PATH="$PASCAL_PREVIEW_PREFIX/bin:$PATH"
 pascal --version
+```
+
+Before activating the preview, save or otherwise persist every project with active work, then disconnect all editor and agent clients from this local service. `pascal update` may stop and restart the managed editor and MCP processes. Keeping the same `PASCAL_HOME` preserves persisted project data in that directory, but it does not preserve unsaved or unbound in-memory changes, undo history, or active client sessions.
+
+```bash
 pascal update --version "$PASCAL_PREVIEW_VERSION"
 pascal editor --no-open
 pascal mcp setup claude # or: pascal mcp setup codex
 ```
 
-The expected archive SHA-256 is `814ffa8c6f6a5fced73bf909c616d9a78feff18fd61fd0b4b7d65e74fad5a33d`. The same-version `update` command installs and activates this CLI's bundled runtime, restarting an older running service when necessary. Keep an existing `PASCAL_HOME` unchanged so stored projects remain in the same data directory; `pascal editor` alone reuses any healthy service, including an older one. Keep the preview prefix on the agent host's `PATH` so its configured `pascal mcp connect` command resolves. This preview is not published on npm.
+The expected archive SHA-256 is `814ffa8c6f6a5fced73bf909c616d9a78feff18fd61fd0b4b7d65e74fad5a33d`. The same-version `update` command installs and activates this CLI's bundled runtime, restarting an older running service when necessary. `pascal editor` alone reuses any healthy service, including an older one, so run `pascal update` when activating the preview. Keep the preview prefix on the agent host's `PATH` so its configured `pascal mcp connect` command resolves. This preview is not published on npm.
 
 Run the setup command for the active host. The MCP command installed in host configuration is `pascal mcp connect`. Local use needs no hosted account and does not upload projects automatically. If the connected MCP schema lacks `check_collisions.candidate`, report the narrower supported result rather than implying the candidate was tested.
 

skills/pascal-3d/references/setup.md

@@ -38,12 +38,17 @@ npm install --global --prefix "$PASCAL_PREVIEW_PREFIX" --ignore-scripts \
   "./pascal-app-cli-${PASCAL_PREVIEW_VERSION}.tgz"
 export PATH="$PASCAL_PREVIEW_PREFIX/bin:$PATH"
 pascal --version
+```
+
+Before activating the preview, save or otherwise persist every project with active work, then disconnect all editor and agent clients from this local service. `pascal update` may stop and restart the managed editor and MCP processes. Keeping the same `PASCAL_HOME` preserves persisted project data in that directory, but it does not preserve unsaved or unbound in-memory changes, undo history, or active client sessions.
+
+```bash
 pascal update --version "$PASCAL_PREVIEW_VERSION"
 pascal editor --no-open
 pascal mcp setup claude # or: pascal mcp setup codex
 ```
 
-The expected archive SHA-256 is `814ffa8c6f6a5fced73bf909c616d9a78feff18fd61fd0b4b7d65e74fad5a33d`. The same-version `update` command installs and activates this CLI's bundled runtime, restarting an older running service when necessary. Keep an existing `PASCAL_HOME` unchanged so stored projects remain in the same data directory; `pascal editor` alone reuses any healthy service, including an older one. Keep the preview prefix on the agent host's `PATH` so its configured `pascal mcp connect` command resolves. This preview is not published on npm.
+The expected archive SHA-256 is `814ffa8c6f6a5fced73bf909c616d9a78feff18fd61fd0b4b7d65e74fad5a33d`. The same-version `update` command installs and activates this CLI's bundled runtime, restarting an older running service when necessary. `pascal editor` alone reuses any healthy service, including an older one, so run `pascal update` when activating the preview. Keep the preview prefix on the agent host's `PATH` so its configured `pascal mcp connect` command resolves. This preview is not published on npm.
 
 Run only the setup command for the active host. For a JSON-based MCP client, use: