epic #470 C14: retire the core-decoupling ratchet; stacked on #808

.github/workflows/ci.yml

@@ -31,72 +31,14 @@ permissions:
 
 jobs:
   # ------------------------------------------------------------------
-  # Core-decoupling check (epic #470). The Dev Platform lives in its own
-  # repository; this fails if core re-acquires ANY reference to it.
-  #
-  # This was a ratchet — a committed count allowed to fall and never rise —
-  # for as long as the extraction was in flight. C13 finished the job and the
-  # floor is now hard zero: the script asserts 0 outright instead of reading
-  # a baseline, because a baseline is a number and numbers get edited. The
-  # job name is kept for continuity as a required check.
+  # The core-decoupling ratchet (epic #470) used to run here. It landed with
+  # the epic on 2026-07-30, peaked at 3,448 counted Dev Platform references,
+  # reached 0 at C13 and was removed at C14 on 2026-08-21. The extraction is
+  # finished — the Dev Platform lives in `byte5ai/omadia-dev-platform` — so
+  # the guard has done its job and a guard with nothing left to guard is
+  # maintenance, not safety. Reintroducing that coupling is now an ordinary
+  # architectural decision to argue for in review, not a number to edit.
   # ------------------------------------------------------------------
-  decoupling:
-    name: core decoupling ratchet (#470)
-    runs-on: ubuntu-latest
-    # Was 5. `Ensure ripgrep` alone took 3:28 on a good run and 5:07 on a bad
-    # one, so this job routinely had ~90 seconds of headroom and intermittently
-    # blew the budget. A timed-out job reports as `cancelled`, and its remaining
-    # steps are `skipped` — so the ratchet was a REQUIRED check that silently
-    # did not run, which is the same failure family as a permanently-green
-    # guard-skip: not red, not passing, just absent.
-    timeout-minutes: 10
-    steps:
-      - uses: actions/checkout@v7
-
-      - uses: actions/setup-node@v7
-        with:
-          node-version: '22'
-
-      - name: Ensure ripgrep
-        # `apt-get update` is the expensive half — it refetches every package
-        # index to install one small binary. Try the cached indices first and
-        # only refresh when that actually fails. `rg` is genuinely needed:
-        # `check-core-decoupling.mjs` shells out to it (`execFileSync('rg', …)`).
-        run: |
-          if ! command -v rg >/dev/null; then
-            sudo apt-get install -y ripgrep \
-              || (sudo apt-get update && sudo apt-get install -y ripgrep)
-          fi
-
-      - name: Assert core has ZERO Dev Platform references
-        # C13 pinned the count at 0 permanently. The script asserts that
-        # outright (`EXTRACTION_COMPLETE`) rather than comparing against
-        # `decoupling-baseline.json`, so a single reintroduced reference —
-        # code path, config key, i18n key, fixture string or comment — fails
-        # this job, and there is no number anyone can edit to make it pass.
-        #
-        # Belt and braces on top of the script's own exit code: assert the
-        # report literally says zero. If a future refactor ever made the
-        # script exit 0 while still counting hits (a `--report`-style path
-        # slipping into the default mode, an early `process.exit(0)`), this
-        # step still fails. A guard that can only fail loudly is the point.
-        run: |
-          node scripts/check-core-decoupling.mjs
-          node scripts/check-core-decoupling.mjs --report | tee /tmp/decoupling.txt
-          grep -qE '^  TOTAL 0( |$)' /tmp/decoupling.txt
-
-      - name: Test the detector itself
-        # The guard above only earns its required-check status if the detector
-        # it runs actually detects. `check-core-decoupling.test.mjs` plants a
-        # probe file under `middleware/src/` and asserts the checker goes red —
-        # it is the mutation test, committed.
-        #
-        # It is wired HERE deliberately. `scripts/` has no package.json and no
-        # other job globs `*.test.mjs`, so a test file sitting there would
-        # never execute: green because nothing ran it, which this job's own
-        # timeout comment above already names as the same failure family as a
-        # permanently-green guard-skip. A test nobody runs is not a test.
-        run: node --test scripts/check-core-decoupling.test.mjs
 
   # ------------------------------------------------------------------
   # Middleware: lint + typecheck + node:test against tsx
@@ -158,7 +100,8 @@ jobs:
       # The root tsconfig only includes src/, so `typecheck` above never sees
       # test/ or scripts/. This ratchet typechecks those trees against a
       # committed baseline: it blocks NEW type errors while the pre-existing
-      # debt burns down (issue #573, same pattern as the decoupling ratchet).
+      # debt burns down (issue #573, same pattern the retired #470
+      # core-decoupling ratchet used).
       - name: Typecheck (test + scripts trees, ratchet)
         run: npm run typecheck:test
 
@@ -253,9 +196,9 @@ jobs:
       # the count drops (broken DSN, a future suite reading an env var we
       # don't set here, a cold-connection flake), a suite silently vanished —
       # go red. Raise PG_TEST_FLOOR when you add Postgres tests, same
-      # baseline-ratchet convention as the core-decoupling job above.
+      # baseline-ratchet convention the retired #470 core-decoupling job used.
       #
-      # NOTE — this floor is the INVERSE of the core-decoupling ratchet, and
+      # NOTE — this floor is the INVERSE of that ratchet, and
       # the reflex from that one is wrong here. It is a MINIMUM on desirable
       # coverage, so raising it as suites land is the intended move; there the
       # counted thing is debt and the rule is reword-don't-raise. It had

docs/CHANGELOG.md

@@ -18,6 +18,23 @@ entry. See `CONTRIBUTING.md` § Releases & changelog.
 
 ## [Unreleased]
 
+### Removed — the core-decoupling ratchet is retired (#470 C14)
+
+- `scripts/check-core-decoupling.mjs`, its colocated detector test
+  `scripts/check-core-decoupling.test.mjs`, `specs/470-dev-platform-plugin/decoupling-baseline.json`
+  and the CI job `core decoupling ratchet (#470)` are removed. The guard ran from 2026-07-30
+  (#539) to 2026-08-21, peaked at **3,448** counted Dev Platform references in core, fell to
+  214 at C10 and reached **0** at C13, where the floor was pinned permanently.
+- It was scaffolding for a migration in flight: it made a file inventory's staleness
+  survivable, because a checklist goes stale on contact and a count does not. The extraction
+  is finished and the Dev Platform lives in `byte5ai/omadia-dev-platform`, so the guard has
+  nothing left to guard — keeping it would cost a CI job and leave an editable number behind.
+  Reintroducing that coupling is now an ordinary architectural decision, argued for in review.
+- No behaviour change: the job is not among `main`'s required status checks and no other job
+  declared `needs: decoupling`, so removing it cannot leave a branch waiting on a check that
+  will never report. (The job's own comment claimed required-check status; branch protection
+  says otherwise — worth knowing before trusting a comment about CI over the API.) The specs
+  under `specs/470-dev-platform-plugin/` keep a closing note in place of the live baseline rows.
 ### Added — runtime install of subscription CLIs from the admin UI (#309 extension, enabler for #294)
 
 - **The Subscription-CLIs page can now install a missing vendor CLI in-app.**

middleware/scripts/check-test-file-durations.mjs

@@ -12,7 +12,8 @@
  *
  * WHY A FRACTION AND NOT A COMMITTED BASELINE
  * -------------------------------------------
- * The repo's other ratchets (core-decoupling, test-typecheck, PG_TEST_FLOOR)
+ * The repo's other ratchets (test-typecheck, PG_TEST_FLOOR, and the #470
+ * core-decoupling ratchet until it was retired at C14)
  * commit an absolute number, because they count things that do not vary by
  * machine. Durations do: the same suite measured 36.5 s locally on 16 cores
  * and 172 s on a 4-vCPU CI runner — a 4.7x spread. An absolute ms baseline

middleware/scripts/check-test-typecheck.mjs

@@ -17,7 +17,8 @@
  *
  * The trees carry a large pre-existing type-error debt (intentional partial
  * mocks, drifted fixtures, a few real bugs). Fixing all of it in one PR is a
- * 130-file diff. So — exactly like `scripts/check-core-decoupling.mjs` — this
+ * 130-file diff. So — exactly like the #470 core-decoupling ratchet did, until
+ * it was retired at C14 — this
  * counts errors per file against a committed baseline and refuses to let any
  * file's count rise. New test type errors fail CI immediately; the debt burns
  * down over time and the baseline only ever falls.

middleware/scripts/plugin-ledger-handoff.mjs

@@ -29,10 +29,9 @@
  *
  * DELIBERATELY GENERIC. It names no plugin and no table: the plan file
  * supplies the plugin id, the ledger, the migrations directory and the
- * entries, and core supplies the donor ledger. That is not tidiness — core's
- * decoupling ratchet (`scripts/check-core-decoupling.mjs`) requires that no
- * core file name the extracted plugin, and the next plugin to leave core will
- * want this tool unchanged.
+ * entries, and core supplies the donor ledger. That is not tidiness — the
+ * extraction rule (epic #470) was that no core file names the extracted
+ * plugin, and the next plugin to leave core will want this tool unchanged.
  *
  * PLAN FILE
  * ---------

middleware/test/auth/staticPublicPathsClosedSet.test.ts

@@ -25,8 +25,8 @@ import { PUBLIC_MCP_PATH } from '../../src/mcp/publicMcpPath.js';
  * WHY THIS SUITE DOES NOT NAME THE TWO DELETED PATHS
  * -------------------------------------------------
  * Naming them would write a plugin's wire paths back into core — the exact
- * coupling this epic removes, and the thing `scripts/check-core-decoupling.mjs`
- * counts on its way to a permanent zero. A test that hardcodes a plugin's URLs
+ * coupling this epic removed, and the thing the #470 decoupling ratchet counted
+ * on its way to zero before it was retired at C14. A test that hardcodes a plugin's URLs
  * is a reference to that plugin whichever direction it asserts in, and it goes
  * stale-but-green the day the plugin renames them.
  *

middleware/test/pluginMigrationHandoff.pg.test.ts

@@ -13,9 +13,9 @@
  *   - `pg_advisory_xact_lock` really serialises two concurrent seeders.
  *   - `ROLLBACK` really un-does the dry run, including the ledger DDL.
  *
- * NAMES ARE DELIBERATELY NEUTRAL. This suite lives in core, and core's
- * decoupling ratchet (`scripts/check-core-decoupling.mjs`) counts every one of
- * the extracted plugin's identifiers in `middleware/test`. The mechanism is
+ * NAMES ARE DELIBERATELY NEUTRAL. This suite lives in core, and the epic #470
+ * extraction rule kept every one of the extracted plugin's identifiers out of
+ * `middleware/test` (the ratchet that enforced it was retired in C14). The mechanism is
  * generic — any plugin adopting any core ledger — so the fixture is generic
  * too. The nine REAL witnesses are documented in the epic README under
  * `specs/470-dev-…-plugin/README.md` and shipped by the plugin repo, which