@adia-ai/adia-ui-forge 0.1.3 → 0.8.0

package/skills/adia-ui-release/assets/case-studies/2026-05-23-freshness-gate-recovery-v0.6.32.md

@@ -1,97 +0,0 @@
-# Case study — v0.6.32 (2026-05-23)
-
-## Freshness-gate recovery: deploy-handoff variant catching pre-flight gap
-
-**Variant:** Mode 3 (Deploy peer handoff) + freshness-gate recovery preamble.
-
-**Status:** Resolved cleanly. 9/9 npm publishes succeeded on first dispatch. Tags re-pointed without force-push.
-
----
-
-## What happened
-
-Peer (Kim) pre-authored v0.6.32 release commit `5285defa8` after a substantive substrate sweep:
-
-- `60a247efa` — tag/badge/segment phantom-gap fix
-- `6db1d8a58` — FB-51 P0: `everything.min.js` now registers all 95+ primitives (was incomplete)
-- `e394a8099` — remove duplicate `data-chunk-kind` attrs
-- `d8f4ff2c0` — admin-shell `[data-col]/[data-row]` CSS scoped to ancestor
-- **`5d8c62a80`** — feat(0.6.32): everything.js bundle fix + icons-manifest + llms.txt + **status:stable sweep on all 141 component YAMLs + anatomy sweep across examples pages**
-- `5285defa8` — release commit + 10 tags created locally
-
-The status:stable yaml sweep + anatomy sweep touched 284 yaml/json files. The chunk corpus is yaml-derived — peer's sweep regenerated the yaml sources but **did not run `harvest:chunks` + `build:embeddings:chunks`** as a follow-on step.
-
-Peer's commit message claimed verification was green with 7 gates enumerated:
-
-> check:lockstep OK / verify:components clean / verify:traits 56/56 / check:css-bundles-fresh 5/5 / check:js-bundles-fresh 6/6 / smoke:engines ok / test:a2ui 22/22
-
-**The remaining 9 of 16 gates were skipped** — including `verify:corpus`, `check:chunks-fresh`, `check:embeddings-fresh`, and `check:links`. The post-yaml drift was invisible to peer's automation.
-
-## What I caught
-
-User instruction: _"re-harvest the chunks then cut 0.6.32"_
-
-Operator probe acting as forensic gate. I picked up the cycle at deploy-handoff state (release commit + tags pre-created locally, tags not yet pushed), ran `npm run harvest:chunks`:
-
-- 387 chunk instances parsed → 238 unique → 190 leaf + 48 nested-expanded
-- **230+ chunk JSON files updated** + `chunks/_index.json` regenerated
-
-Then `check:embeddings-fresh` reported:
-
-```text
-[embeddings] FAILED — drift detected:
-  ✗ chunk-embeddings.json is 1.0d older than chunks/_index.json
-    — run `npm run build:embeddings:chunks`
-```
-
-## Recovery (clean — no force-push)
-
-Because peer's tags were local-only (never pushed to origin), recovery was a tag-move, not a tag-rewrite:
-
-1. `npm run build:embeddings:chunks` — regenerated chunk-embeddings.json (OpenAI text-embedding-3-small, 8 batches, 238 × 1536d, ~6.8 MB)
-2. Updated `@adia-ai/a2ui-corpus [0.6.32]` CHANGELOG entry to document the regen (replaced peer's `_No pending changes._` placeholder).
-3. Committed `chunks/` + `chunk-embeddings.json` + CHANGELOG as `e864fc396 chore(a2ui-corpus): v0.6.32 re-harvest — 238 chunks + embeddings regen`.
-4. Re-ran the FULL 16-gate Step 3 roster. All green (`verify:corpus` exit code 0; 65 pre-existing warnings carried forward — 61 field-wraps-widget + 4 enum-out-of-range, all technical debt unrelated to v0.6.32).
-5. Deleted local tags at `5285defa8`, recreated all 10 at `e864fc396`:
-
-   ```bash
-   for tag in v0.6.32 web-{components,modules}-v0.6.32 llm-v0.6.32 \
-              a2ui-{compose,corpus,mcp,retrieval,runtime,validator}-v0.6.32; do
-     git tag -d $tag && git tag $tag HEAD
-   done
-   ```
-
-6. F-N1 (`check:release --all-pending`): 9/9 per-package clean.
-7. Pushed branch + 10 tags, dispatched 9 publish workflows → 9/9 success on first dispatch.
-
-## Key lessons graduated into the skill
-
-### 1. Step 3 grew §Step 3.0 (harvest preamble)
-
-When source content changes (yaml / examples / contents.html with data-chunk-\* annotations / status:stable sweeps / anatomy sweeps), proactively run `harvest:chunks` + `build:embeddings:chunks` BEFORE the gates — not reactively after they fail. Cheap to run, catches drift before it lands in a tag.
-
-Documented in `cycle-happy-path.md` §Step 3.0 with source-content signal list + recipe.
-
-### 2. Step 3 grew an explicit "all 16 must run" rule
-
-The previous Step 3 listed the gates but didn't forbid selective subsets. Peer's automation enumerated 7. The rule is now:
-
-> **All 16 gates MUST run. Selective subset = pre-flight failure.**
-
-Documented in `cycle-happy-path.md` §Step 3.1 with the canonical roster numbered 1-16.
-
-### 3. The recovery recipe is preserved as this case study
-
-Future cycles facing the same situation can follow this playbook:
-
-- Tags local-only → safe to delete + recreate at new HEAD
-- Tags pushed → either force-push tags (rare, document why) or roll forward to next version
-- Always: re-run F-N1 after tag-move; full 16-gate roster after harvest preamble.
-
-## Cross-references
-
-- Audit-history ledger: `.brain/audit-history/2026-05-23-release-cut-v0.6.32.json`
-- Release notes draft: `/tmp/release-v0.6.32/notes.md`
-- Sibling case studies:
-  - [corpus-drift-remediation-v0.6.15](./2026-05-20-corpus-drift-remediation-v0.6.15.md) — earlier embeddings drift (6.6 days stale, batch-push context).
-  - [version-skip-correction-v0.6.12](./2026-05-20-version-skip-correction-v0.6.12.md) — adjacent recovery pattern (tag-rewrite, force-push variant).

package/skills/adia-ui-release/assets/case-studies/2026-05-26-catalog-drift-recurring-v0.6.40.md

@@ -1,147 +0,0 @@
-# Case study — v0.6.37 + v0.6.40 (2026-05-25 / 2026-05-26)
-
-## Catalog-drift recurring — §Step 3.0 graduation was incomplete; v1.9.0 closes the output-staging gap
-
-**Variant:** Mode 1 (Cut & ship) — both occurrences. Same class of CI failure (catalog drift on `verify:components`), same recovery shape (regen + commit follow-up + force-move 10 tags + re-dispatch).
-
-**Status:** Both cycles published 9/9 net of recovery. v0.6.37 took ~5 min of recovery; v0.6.40 took ~5 min of recovery on the **same** class. v1.9.0 graduates the gap that v1.8.0 missed.
-
----
-
-## What happened — twice
-
-### First occurrence — v0.6.37 (2026-05-25)
-
-7/9 publish workflows failed at the `verify:components` step. The diagnostic was:
-
-```text
-[verify] DRIFT catalog: packages/a2ui/corpus/catalog-a2ui_0_9.json
-[verify] DRIFT rules: packages/a2ui/corpus/catalog-a2ui_0_9_rules.txt
-[verify] 2 file(s) would change. Run build-components.mjs to update.
-##[error]Process completed with exit code 1.
-```
-
-Root cause: catalog files were modified in WT by peer (uncommitted local audit-token-pair or yaml-sweep work). The release agent correctly excluded the WT pre-state from the release-commit allowlist per multi-agent baseline discipline. **§Step 3.0 didn't run a catalog regen** — the v1.7.0 graduation covered `harvest:chunks` + `build:embeddings:chunks` but not `node scripts/build/components.mjs` (catalog regen).
-
-Recovery: regen + commit follow-up + force-move 10 tags + re-dispatch.
-
-### v1.8.0 graduation — incomplete
-
-The v0.6.37 case study lesson graduated into **adia-ui-release v1.8.0**, which added `node scripts/build/components.mjs` (catalog regen) to §Step 3.0 alongside `harvest:chunks` + `build:embeddings:chunks` + `build:bundles`.
-
-**The graduation covered REGEN but not OUTPUT-STAGING.** Step 3.0 told the agent to regenerate the catalog. It didn't tell the agent that the regen output had to LAND in the release commit even when the same file paths were also peer-in-flight in the working tree.
-
-### Second occurrence — v0.6.40 (2026-05-26)
-
-8/9 publish workflows failed at `verify:components` — only `llm` succeeded (pure ride-along, no catalog dependency). Same diagnostic; same root cause shape.
-
-The release agent followed v1.8.0 §Step 3.0 to the letter — ran `node scripts/build/components.mjs`, got fresh catalog output, ran the gates locally (`verify:components` clean). Then inspected WT:
-
-```text
- M packages/a2ui/corpus/catalog-a2ui_0_9.json       ← peer-in-flight + my regen overlay
- M packages/a2ui/corpus/catalog-a2ui_0_9_rules.txt  ← same
- M packages/web-components/components/badge/badge.d.ts      ← peer-in-flight (NOT mine)
- M packages/web-components/components/card/card.css         ← peer-in-flight (NOT mine)
- M packages/web-components/components/progress-row/progress-row.css  ← peer-in-flight
-```
-
-Multi-agent baseline discipline says "stage explicit allowlist; never `git add -A`." Catalog files were on the WT pre-state list as peer-in-flight, so the agent excluded them. **The regen output (correct for this release) also got excluded — same file paths.**
-
-Tags pushed. CI hit DRIFT. 8/9 failed.
-
-Recovery: identical to v0.6.37 — regen + commit follow-up + force-move 10 tags + re-dispatch. ~5 min.
-
-**Twice in the same class is a missing trip-wire.**
-
----
-
-## What I caught (post-mortem clarity)
-
-The v1.8.0 wording said _regenerate the catalog_. It didn't say:
-
-- "The regen output supersedes any peer-in-flight WT pre-state on the same paths."
-- "Catalog files are non-negotiable in the release-commit allowlist when source content changed in the window."
-- "Re-verify with `verify:components` after staging — if DRIFT, you excluded the regen output."
-
-All three would have caught the v0.6.40 occurrence. The conflict between "exclude peer WT pre-state" (baseline discipline) and "stage regen output" (§Step 3.0) had no documented winner. The agent resolved the conflict toward exclusion both times.
-
----
-
-## v1.9.0 graduation — three layers, defense in depth
-
-### Layer 1 — Stronger §Step 3.0 wording (supersedes-WT contract)
-
-The regen step has FOUR deliverables (was implicit; now enumerated):
-
-```bash
-node scripts/build/components.mjs           # → catalog + sidecar
-npm run harvest:chunks                       # → chunks/_index + per-chunk JSON
-npm run build:embeddings:chunks              # → chunk-embeddings.json
-npm run build:bundles                        # → dist/*.min.{css,js}
-```
-
-**The regen output supersedes any peer-in-flight WT pre-state on the same file paths.** Peer's uncommitted catalog modifications are superseded by the fresh regen — stage the regenerated catalog into the release-commit allowlist UNCONDITIONALLY. Peer can rebase any divergent work on top of the release commit after the cycle.
-
-### Layer 2 — NEW §Step 5.5 — pre-commit verification trip-wire
-
-After `git add` (Step 5) but BEFORE `git commit`, re-run the verify gates one final time:
-
-```bash
-node scripts/build/components.mjs --verify    # catalog drift?
-npm run check:chunks-fresh
-npm run check:embeddings-fresh
-```
-
-**Any DRIFT here means a regen output was excluded from staging.** Recovery is in-cycle (no force-push): `git add <drift-paths>` → re-run `verify` until clean → Step 6 (commit).
-
-The trip-wire runs in <2 seconds locally vs ~5 minutes to recover from a CI failure. Pure win.
-
-### Layer 3 — §Step 5 allowlist convention update
-
-When source content changed in this window (yaml / examples.html / contents.html with `data-chunk-*` / status:stable sweeps / Wave commits / OD-5 sweeps), the release-commit allowlist MUST include these regen-deterministic outputs UNCONDITIONALLY:
-
-- `packages/a2ui/corpus/catalog-a2ui_0_9.json`
-- `packages/a2ui/corpus/catalog-a2ui_0_9_rules.txt`
-- `packages/a2ui/corpus/chunk-embeddings.json`
-- `packages/a2ui/corpus/chunks/` (if chunk content drifted)
-- `packages/web-components/dist/web-components.min.{css,js}`
-- `packages/web-components/dist/icons-manifest.js`
-- `packages/web-modules/dist/everything.min.js`
-- `packages/web-modules/dist/shell/admin-shell.min.js`
-- `packages/web-modules/dist/icons-manifest.js`
-
-These are regen-deterministic outputs; their WT pre-state (if any) is always superseded by the release-cycle regen.
-
-This is **not** "stage WT changes" (still violates multi-agent baseline). This is "stage the regen outputs you produced in §Step 3.0 — they supersede any peer WT pre-state on the same paths."
-
----
-
-## How v1.9.0 closes the gap (verification target for the next cycle)
-
-For the next v0.6.41+ cycle:
-
-1. Verify peer-in-flight catalog modifications exist in WT (simulate by `touch packages/a2ui/corpus/catalog-a2ui_0_9.json` if needed).
-2. Run §Step 3.0 (catalog regen).
-3. Stage allowlist per the new convention — catalog UNCONDITIONALLY included.
-4. Run §Step 5.5 trip-wire (`node scripts/build/components.mjs --verify` + `check:chunks-fresh` + `check:embeddings-fresh`).
-5. Expected: clean. Commit lands; CI passes 9/9 on first dispatch.
-
-If the trip-wire fails: the v1.9.0 graduation is still incomplete; back to the bench.
-
----
-
-## Cross-references
-
-- Sibling case studies:
-  - [freshness-gate-recovery-v0.6.32](./2026-05-23-freshness-gate-recovery-v0.6.32.md) — same pattern shape: CI failure detected at publish time → graduation into §Step 3.0 (this case is the v1.8.0 → v1.9.0 iteration). v0.6.32 → v1.7.0 graduated cleanly; v0.6.37 → v1.8.0 graduated incompletely.
-  - [corpus-drift-remediation-v0.6.15](./2026-05-20-corpus-drift-remediation-v0.6.15.md) — earlier embeddings drift (6.6 days stale, batch-push context).
-- Audit-history ledgers:
-  - `.brain/audit-history/2026-05-25-release-cut-v0.6.37.json`
-  - `.brain/audit-history/2026-05-26-release-cut-v0.6.40.json`
-- FEEDBACK ticket that motivated this graduation: `FEEDBACK-75` (the internal release-cycle self-feedback ticket)
-
-## Lesson aphorism
-
-> "Once is happenstance, twice is coincidence, twice in the same class is a missing trip-wire."
-
-The release skill survives because every CI failure that recurs gets graduated into mechanical defense, not into a longer recovery runbook.

package/skills/adia-ui-release/assets/templates/stub-changelog.template.md

@@ -1,22 +0,0 @@
-<!--
-Lockstep stub block — for a ride-along package with no source changes
-in this release. The substantive description points at the relevant
-@adia-ai/* package whose CHANGELOG carries the headline.
-
-Variables (replace before use):
-  $VERSION       — the version being cut (e.g. 0.6.21)
-  $DATE          — release date in YYYY-MM-DD
-  $SUBSTANTIVE   — one-sentence summary of the substantive work and
-                   where it shipped, with the cross-reference path.
-
-Example invocation:
-  node scripts/insert-stub.mjs --version 0.6.21 --date 2026-05-21 \
-    --substantive "table-ui truncate-default (§403) and FEEDBACK-37 retraction in @adia-ai/web-modules" \
-    --xref "packages/web-modules/CHANGELOG.md#0621--2026-05-21"
--->
-
-## [$VERSION] — $DATE
-
-### Maintenance
-
-- **Lockstep version bump only.** No source changes in this package; bumped to maintain the 9-package version coherence enforced by `scripts/release/check-lockstep.mjs`. Substantive v$VERSION work shipped in $SUBSTANTIVE. See `$XREF` for details.

package/skills/adia-ui-release/evals/evals.json

@@ -1,164 +0,0 @@
-{
-  "$schema": "https://anthropic.com/schemas/skill-evals.v1.json",
-  "skill": "adia-ui-release",
-  "version": "0.1.0",
-  "description": "Mode-trigger evals + diagnostic-routing evals for the autonomous release engineer (release cycle + migration-guide authoring).",
-  "evals": [
-    {
-      "id": "mode-1-bare-cut",
-      "category": "mode-trigger",
-      "input": "prepare and initiate 0.6.22 release",
-      "expected_mode": 1,
-      "expected_loads": ["cycle-happy-path.md", "multi-agent-baseline.md"],
-      "must_NOT_load": ["rollup-notes.md", "recovery-paths.md"]
-    },
-    {
-      "id": "mode-2-from-scratch",
-      "category": "mode-trigger",
-      "input": "the peer landed feedback fixes under [Unreleased] — cut v0.6.22 from scratch",
-      "expected_mode": 2,
-      "expected_loads": ["cycle-happy-path.md", "changelog-discipline.md"],
-      "must_NOT_load": ["rollup-notes.md"]
-    },
-    {
-      "id": "mode-3-deploy-handoff",
-      "category": "mode-trigger",
-      "input": "peer pre-cut the v0.6.22 release commit; deploy + publish it",
-      "expected_mode": 3,
-      "expected_loads": ["cycle-happy-path.md"]
-    },
-    {
-      "id": "mode-4-batch-push",
-      "category": "mode-trigger",
-      "input": "there are 2 unpushed release commits — ship them together",
-      "expected_mode": 4,
-      "expected_loads": ["recovery-paths.md"]
-    },
-    {
-      "id": "mode-5-rollup",
-      "category": "mode-trigger",
-      "input": "write a consolidated rollup release note for v0.6.16 → v0.6.21",
-      "expected_mode": 5,
-      "expected_loads": ["rollup-notes.md"],
-      "must_NOT_load": ["cycle-happy-path.md", "multi-agent-baseline.md"]
-    },
-    {
-      "id": "mode-6-just-verify",
-      "category": "mode-trigger",
-      "input": "run the pre-flight gates — I just want to smoke-check",
-      "expected_mode": 6,
-      "expected_loads": ["gates-catalog.md"],
-      "must_NOT_load": ["cycle-happy-path.md", "notes-authoring.md"]
-    },
-    {
-      "id": "mode-7-recovery",
-      "category": "mode-trigger",
-      "input": "we just shipped v0.6.22 and a consumer reports the demo-shell is broken",
-      "expected_mode": 7,
-      "expected_loads": ["recovery-paths.md"],
-      "may_route_to": ["ops-postmortem"]
-    },
-    {
-      "id": "mode-8-investigation",
-      "category": "mode-trigger",
-      "input": "what shipped in v0.6.18? what did we miss?",
-      "expected_mode": 8,
-      "expected_loads": ["ledger-discipline.md"]
-    },
-    {
-      "id": "mode-9-notes-only",
-      "category": "mode-trigger",
-      "input": "I just need the GH release body for v0.6.22 — don't cut anything",
-      "expected_mode": 9,
-      "expected_loads": ["notes-authoring.md"],
-      "must_NOT_load": ["cycle-happy-path.md", "multi-agent-baseline.md"]
-    },
-    {
-      "id": "mode-11-migration-guide-authoring",
-      "category": "mode-trigger",
-      "input": "we're cutting a breaking MINOR that renames variant=danger to color=danger — author the MIGRATION GUIDE section",
-      "expected_mode": 11,
-      "expected_loads": ["migration-guide-authoring.md"],
-      "must_NOT_load": ["rollup-notes.md"]
-    },
-    {
-      "id": "diagnostic-migration-consumer-side-rejected",
-      "category": "diagnostic-routing",
-      "input": "migrate our downstream consumer app to @adia-ai/web-components 0.0.20",
-      "expected_route": "NOT this skill — consumer-side migration lives in the separate consumer/app-author plugin; this skill owns only the PRODUCER side (authoring the guide + sweeping in-repo surfaces when cutting the breaking release)",
-      "expected_action": "decline + redirect to the consumer/app-author plugin"
-    },
-    {
-      "id": "bare-activation",
-      "category": "cold-start",
-      "input": "use adia-ui-release",
-      "expected_behavior": "render the mode triage menu verbatim + ask for mode+version; do NOT load any references/*.md bundle",
-      "expected_loads": []
-    },
-    {
-      "id": "diagnostic-fn1-warn",
-      "category": "diagnostic-routing",
-      "input": "F-N1 says: '[warn] diff packages/web-components/components/ touched but CHANGELOG [0.6.22] doesn't mention components'",
-      "expected_route": "changelog-discipline.md § F-N1 diff-coverage enrichment",
-      "expected_action": "enrich an existing CHANGELOG entry to include the 'components/' path keyword; amend release commit; delete + re-tag; re-run F-N1",
-      "expected_severity": "cosmetic"
-    },
-    {
-      "id": "diagnostic-demo-shells-fail",
-      "category": "diagnostic-routing",
-      "input": "check:demo-shells fails: 'demo packages/web-components/components/table/table.html missing skeleton-ui'",
-      "expected_route": "gates-catalog.md § Category 3 § check:demo-shells",
-      "expected_action": "add <script type='module' src='../skeleton/skeleton.js'> to the demo .html",
-      "expected_severity": "HIGH (release-blocking)"
-    },
-    {
-      "id": "diagnostic-embeddings-stale",
-      "category": "diagnostic-routing",
-      "input": "check:embeddings-fresh FAILED — chunk-embeddings.json is 6.6d older than chunks/_index.json",
-      "expected_route": "gates-catalog.md § Category 7 § check:embeddings-fresh",
-      "expected_action": "npm run build:embeddings:chunks; stage chunk-embeddings.json; add a CHANGELOG note to a2ui-corpus [vX.Y.Z]"
-    },
-    {
-      "id": "diagnostic-version-skip",
-      "category": "diagnostic-routing",
-      "input": "peer's release commit bumped 0.6.11 → 0.6.13 but I want to cut 0.6.12",
-      "expected_route": "recovery-paths.md § Scenario 1 (version-skip correction)",
-      "expected_action": "correction commit on top of peer's commit; rewrite v0.6.13 → v0.6.12 across pkg.json + CHANGELOGs + CSS comments + ADRs + tickets + lockfile; tag v0.6.12 at correction commit",
-      "case_study": "assets/case-studies/2026-05-20-version-skip-correction-v0.6.12.md"
-    },
-    {
-      "id": "diagnostic-stale-test",
-      "category": "diagnostic-routing",
-      "input": "vitest fails: admin-shell.test.js expects '[slot=heading] { display: none }' but the source has ':is(...)[slot=heading], [slot=heading]:not(:has(> [slot])) { display: none }'",
-      "expected_route": "recovery-paths.md § Scenario 5 (stale test detection)",
-      "expected_action": "read the CHANGELOG [vX.Y.Z] entry — does it describe the new behavior? If yes → update the test assertion to match the new shape, add a comment naming the cycle, add a guard assertion for the new contract. If no → real regression, fix the source."
-    },
-    {
-      "id": "diagnostic-peer-in-flight-stray",
-      "category": "diagnostic-routing",
-      "input": "working tree has admin-shell.tokens.css modified (uncommitted, undocumented, contradicts the [VERSION] CHANGELOG entry)",
-      "expected_route": "multi-agent-baseline.md § Discipline 4 (stash workflow)",
-      "expected_action": "git stash push the file with a clear message; ship without; restore after; flag to operator; note in ledger"
-    },
-    {
-      "id": "stash-pop-conflict-after-release",
-      "category": "diagnostic-routing",
-      "input": "after the cycle, git stash pop reports 'kept the stash' — concurrent peer activity",
-      "expected_route": "recovery-paths.md § Scenario 6 (concurrent peer mid-cycle)",
-      "expected_action": "compare stash content vs current HEAD; if peer committed equivalent content, drop stash (redundant); if not, leave stashed; document in ledger"
-    },
-    {
-      "id": "behavior-shipped-needs-retract",
-      "category": "diagnostic-routing",
-      "input": "FEEDBACK-N shipped in vN.M.X but is now diagnosed as misdiagnosis — how to retract?",
-      "expected_route": "case-studies/2026-05-21-feedback37-retraction-v0.6.21.md",
-      "expected_action": "frame retraction as a 'Fixed — no longer emits spurious X' entry in the next cycle's CHANGELOG; add a heads-up paragraph in the release notes; consider ADR if the retraction reveals a framing issue worth ratifying"
-    }
-  ],
-  "non_eval_invariants": [
-    "Bare activation MUST NOT exceed 500 lines loaded (SKILL.md + skill.json + CHANGELOG.md only).",
-    "Modes 5, 6, 9, 11 (single-purpose modes) each MUST NOT load cycle-happy-path.md or multi-agent-baseline.md.",
-    "Every case-study referenced from an eval MUST exist at the named path.",
-    "Every CLI helper named in cycle-happy-path.md MUST exist at the named path."
-  ]
-}