fix(objectql,runtime,rest): store a serializable manifest projection in the package registry - #14499

Fixes #14309

SchemaRegistry.installPackage now stores a serializable projection of the
manifest instead of the caller's live defineStack() object, and the /packages
read doors project a registry entry onto its declared record fields instead of
spreading it whole.

The measured defect

Reproduced on this branch's base, showcase on a random high port
(objectstack dev --seed-admin --fresh -p 39117), signed in as the seeded admin:

Error body of the first two:

Converting circular structure to JSON
    --> starting at object with constructor '_ObjectQL'
    |     property 'actionActivation' -> object with constructor 'ActionActivationProjection'
    |     property 'store' -> object with constructor 'ObjectStoreActionActivationStore'
    --- property 'engine' closes the circle

Which key actually reached the engine

Rather than reason about it, the running server was instrumented at
SchemaRegistry.getPackage (temporary, env-gated, reverted — the file is
byte-identical to its committed state, git hash-object verified) to classify
every top-level manifest key of every installed package by whether it
survives JSON.stringify. Across the 26 packages a showcase boot installs,
exactly one key on one package failed:

[OS_DEBUG_PKG_SERIALIZE] id=com.example.showcase keys=[…40 keys…]
  UNSERIALIZABLE=["plugins :: Array<ConnectorOpenApiPlugin|ConnectorMcpPlugin|ConnectorRestPlugin|
   ConnectorSlackPlugin|MarketplaceProxyPlugin|MarketplaceInstallLocalPlugin|CloudConnectionPlugin|
   RuntimeConfigPlugin> :: Converting circular structure to JSON"]

And only after plugin init. The same manifest was probed 40 times during
boot with UNSERIALIZABLE=[] and 381 times afterwards with the plugins
failure — the transition is a single clean split in the log. That rules out a
whole family of "fixes": an install-time JSON.stringify probe, or any check
that runs while the record is being written, would have called this record
healthy. The projection therefore drops by shape, not by trial
serialisation, and does not depend on when it is asked.

The repair, at the producer

installPackage stored the argument verbatim as pkg.manifest. The registry
item is a record, not the runtime, so it now holds only the declarative
half. The rule is structural rather than a plugins denylist, because the fault
is not "the key is called plugins" — it is "a live object reached the record":

• primitives, plain objects, arrays and Date are data and survive;
• functions / symbols / bigints are dropped (a function was already invisible to
  JSON.stringify; a bigint would have thrown);
• class instances, Map, Set and other exotic objects are dropped;
• a reference cycle among plain data is dropped at the back-edge, so a
  self-referencing manifest degrades to a missing field instead of throwing.

Nothing downstream loses a member it was using: the kernel keeps the live object
in ObjectQL.manifests, and the only reader of manifest.plugins[] —
registerApp's nested-plugin seam — reads its own parameter, never the
record (grepped, and pinned). The caller's manifest is copied, never stripped in
place, which the pins assert both ways.

⛔ Not a consumer-side tolerance (AGENTS.md Prime Directive #12): no reader is
taught to survive an unserializable record — the record is never unserializable.

Premise correction: which handler answered the 500

The card attributes the list/detail 500 to packages/rest/src/package-routes.ts
{ ...item, source: 'registry' }. Measured, that is not the handler that
answered in the showcase composition — packages/runtime/src/domains/packages.ts
did. Two independent probes agree:

• the 404 wording. GET /api/v1/packages/no.such.package returned
  Package 'no.such.package' not found (the runtime domain's spelling); the
  REST twin says Package "no.such.package" was not found.
• the response shape. GET /api/v1/packages/com.objectstack.setup returned the
  installed-package record directly under data (deps.success(pkg)); the
  REST twin wraps it as data.package and stamps source.

So the route-level defence is applied at both twins, and
packages/runtime/src/domains/packages.ts is a declared addition to this card's
file surface (announced on the issue before the work landed). Applying it only
where the card pointed would have shipped a defence with a hole exactly where the
defect was observed.

The doors now project onto the declared InstalledPackageSchema fields, so an
undeclared member appearing on the registry item — a live handle, a
back-reference — degrades to a field the response never mentions, instead of one
bad member on one package failing the whole list for every caller. The REST
merge's database half is deliberately left spreading: its shape belongs to
PackageService, and narrowing it would be a contract decision this card does
not carry. On the REST side _diagnostics is carried explicitly, because
decorateMetadataItem grafts it onto every item leaving getMetaItems and it is
measured to be the only thing that decoration adds for type: 'package'.

No response field is added or renamed; responses that already served fine are
byte-identical.

Verification

git rev-parse --short HEAD at the time of the runs below: 5b8b76e06.

Live routes, same showcase boot, after the fix — the table above; GET /packages
returns all 26 packages, the showcase manifest keeps its 40 declarative keys
(24 objects, 29 flows, 1 app), plugins serves as [] and the function-valued
onEnable is gone. JSON.stringify of the whole listing succeeds.

Package suites (all green, on the commit named above):

@objectstack/objectql   Test Files  255 passed (255)   Tests  4400 passed (4400)
@objectstack/rest       Test Files  168 passed (168)   Tests  2808 passed (2808)
@objectstack/runtime    Test Files  208 passed (208)   Tests  3066 passed (3066)

New pins — 20 cases over three files:
packages/objectql/src/registry-package-manifest-serializable.test.ts (9),
packages/runtime/src/domains/packages-serializable-response.test.ts (5),
packages/rest/src/package-registry-item-projection.test.ts (6).

Typecheck: @objectstack/objectql, @objectstack/runtime, @objectstack/rest
all exit 0. ⚠️ Measured rather than assumed: each package's tsc --noEmit
program excludes **/*.test.ts, so --listFiles reports 0 hits for all three
new test files there. The test layer is covered separately — objectql and rest
run check:test-typecheck over tsconfig.test.json, whose --listFiles does
name the new files (1 hit each, 0 errors attributed to them). packages/runtime
has no test-typecheck wiring at all, so its new test file is covered by no gate;
it was type-checked standalone (tsc --noEmit --ignoreConfig --strict, exit 0, 0
errors) and that gap is reported to the PM rather than papered over.

Gate families — 62 derived by node scripts/pm/dispatch-gates.mjs --commands
(no paths passed; it takes its own change set from the merge base). 60 ran green
locally. The remaining two exited 3, which each script's own verdict text
declares is NOT MEASURED and explicitly "not a red": check-test-completeness
needs a saved turbo run test log it was not handed, and
scripts/pm/check-half-states.mjs needs a per-card comment fetch that never
happened.

check:system-context-census was the one family this diff genuinely moved: the
+50-line helper in packages/runtime/src/domains/packages.ts shifted three
elevation-read anchors (:95 -> :145, :128 -> :178, :246 -> :296). Repaired
mechanically with the gate's own --fix, which re-anchored
content/docs/permissions/system-context.mdx; the gate then reports
OK — 109 elevation read sites in 20 packages across 45 files, all anchored.
That docs edit pulled 24 further doc families into the derivation, all of which
were run and are green.

Ablation (the pin fails without the fix). manifest: toRecordManifest(manifest)
was reverted to manifest, in installPackage; the mutation was confirmed on
disk before running (anchor text 1 -> 0, injected text 1, blob hash differing
from HEAD), the restore leg was git checkout HEAD -- <abs path> under an
EXIT/INT/TERM trap, and restoration was proved by an empty git diff HEAD plus
git hash-object equality with the HEAD blob. No rebuild was needed: the pin
imports ./registry from source, not through dist. Result:

Tests  7 failed | 2 passed (9)     ← ablated
Tests  9 passed (9)                ← restored
AssertionError: expected [Function] to not throw an error but
  'TypeError: Converting circular struct…' was thrown

Declared narrowing — verification ran UNLOCKED. scripts/pm/os-verify-lock.sh
could not take the shared verify lock on this host: no usable flock. The shared
verify lock is declared Linux-only (flock is util-linux, and a stock macOS does
not ship it), so the commands were run directly, without the lock — a declared
narrowing, not a silent one. No serialization guarantee held for these runs, nor
for any sibling agent in this container while they ran.

Declared narrowing — repo-wide scans left to CI. pnpm lint and the rest of
the 62-family farm were not run locally; CI runs the farm exactly once on this
PR. What ran locally is the targeted set and the 60 gate families
named above.

Generated by Claude Code