Spec #44: architecture program — integration branch

[diegohb]

Closes #44, #20, #45, #46, #47, #48, #49, #50.

Summary

The suite's install machinery existed twice — once in src/, once vendored into every generated package via templates — with nothing asserting the two agreed. This program carves the shared machinery into one deep core module, makes the template↔suite relationship a classified, tested table, and replaces the ~20-minute mechanical scaffolding agent pass with a CLI.

src/
├── core/                    # the deep module — exported at "./core" (ADR-0013)
│   ├── entry-predicate.ts   # one classifier: does this entry name our package
│   ├── config-reader.ts     # entries(scope, projectDir) — both jsonc strategies
│   ├── config-splicer.ts    # surgical splice (offset-preserving)
│   ├── registration-detector.ts
│   ├── manifest.ts          # io + sha256 hashes (generated format, frozen)
│   ├── environment.ts       # the only place process.env/console/fetch are read
│   ├── status-reporter.ts   # effective-version pipeline
│   └── …                    # plugin-entry, plugin-name, npm-cache, scope-base
├── scaffold/
│   ├── renderer.ts          # templates → package tree (ex-test-harness renderer)
│   ├── scaffolder.ts        # fresh mode: flags/questions → conformant tree
│   └── promoter.ts          # promote mode: copy → ensure reference → verify → consent-gated retire
└── …                        # suite adapter — consumes core at the seam
templates/                    # adapter renderings over opencode-architect/core, no longer vendored machinery

 generated package install
-  vendored src/installer.ts, config editor, name matcher, manifest io
+  import { … } from "opencode-architect/core"
+
 sync question "are templates and suite in sync?"
-  read files two at a time, hope
+  references/template-sync-table.md — byte-synced | declared-divergent | fixture-covered, all test-enforced
+
 package creation
-  ~20 min of agent file generation (source of every field deviation)
+  bunx opencode-architect scaffold --name … / --promote <path>

Policy is declared at the seam, not parameterized: each adapter (suite CLI, generated CLI) carries named constants for its scope coverage, cache modes, and status flags; the divergence manifest is the executable record. ADR-0013 records the design-twice comparison and the frozen manifest-format compatibility rule. ADR-0006 invariants and ADR-0004 mode exclusivity are preserved; hand-edited installed files keep consumer-modification protection.

Evidence

• Before: main — 276 tests; template sync asserted for 2 pairs; no scaffold; config-entries.ts + monolithic editor; env mutated/console.warn intercepted in 4 test suites.
• After: 313 pass / 0 fail (1250 assertions, 23 files, ~12s), tsc --noEmit clean:

Ran 313 tests across 23 files. 313 pass, 0 fail.

Rendered-fixture proof (end-to-end, generated package depends on opencode-architect/core via a node_modules junction — no registry access): typecheck through the dependency, its own test gate, v2 module contract, install --mode plugin (config + manifest asserted), status (Local: mode=plugin, should-load verdict), load-hook activation with payload hashes, uninstall — all green (tests/templates.test.ts, tests/scaffold.test.ts, tests/core.test.ts).

Merge Danger

Door: two-way

Rollback is a branch revert; no data migration — the manifest compatibility rule makes adopting (or un-adopting) the core dependency a non-event for existing manifests, and no real consumer package has been migrated (promote is the per-package, owner-driven mechanism).

Blast Radius: suite-wide

Every install/status/uninstall path on both adapters now routes through src/core; a core defect reaches all generated packages at once (that is the point — fix once, reach everywhere). Generated packages gain a file:/registry dependency on this package; consumer-modification semantics, skills-market discoverability/install parity, and copy/plugin mode exclusivity are unchanged and test-pinned.