Skillscript

Discovery surface fix. Closes the structural-half of v0.16.2's P0 (the
routing fix shipped without the discovery half: runtime_capabilities
returned byte-identical entries for distinct LocalModel instances, so
adopters had no way to inspect which underlying model each alias was
bound to). Empirically scoped via warm-adopter audit: the discovery-
opacity pattern was uniform across every substrate slot, not just
LocalModel. SkillStore/DataStore/LocalModel/McpConnector all exposed only
their static capabilities; the dynamic manifest() payload — endpoint,
default_model, supported_filters, root_dir, tools_available, etc. — was
unreadable from runtime_capabilities.

Fix: runtime_capabilities now folds each registered connector
instance's manifest() payload into its entry, alongside the existing
features field. One structural fix, applied uniformly across all
slots.

Three observable states per entry:

1. Working — manifest: {capabilities_version, manifest:{...}}
   probed via instance.manifest().
2. Runtime failure — manifest: null, manifest_error: "<message>"
   when instance.manifest() throws (substrate unreachable, parse
   error, etc.). Distinct from #3 so dashboards can differentiate
   "instance broken, ping operator" from "by design".
3. Structural absence — manifest: null, manifest_unsupported: true
   for AgentConnector entries. The v0.9.6 audit excluded manifest()
   from the AgentConnector contract; the type system enforces this
   (ManifestInfo is Exclude<ConnectorType, "agent_connector">).
   Separate sentinel preserves the structural-absence vs. runtime-
   failure distinction.

Bridge wraps convention rides through. The bundled bridge connectors
(LocalModelMcpConnector, DataStoreMcpConnector, SkillStoreMcpConnector)
already implement a manifest().wraps convention that re-exposes the
underlying substrate's full manifest. Capabilities now reads this — a
bridge entry's manifest.manifest.wraps carries the full inner-substrate
manifest. Adopters writing custom bridges get this automatically by
returning {kind, wraps: await innerSubstrate.manifest()} from their
manifest().

Why this scope vs. a LocalModel-only patch

The warm-adopter audit (bfd776a9) ran the manifest-vs-capabilities
diff across all four supported substrate slots and found the same
discovery-opacity pattern in every one. Patching LocalModel only would
have left three substrate slots opaque + introduced a "fix-one-then-
retrofit-three" pattern. The structural fix closes all of them in one
pass.

Internal

• describeEntry in src/mcp-server.ts:885 becomes async + probes
  instance.manifest() per entry with try/catch error handling. Five
  call sites in runtimeCapabilities() lines 837-857 awaited.
• manifest/manifest_error/manifest_unsupported typed in new
  DescribeEntryResult return shape; AgentConnector entries get the
  manifest_unsupported:true sentinel via the runtime check on
  entry.instance.manifest not being a function (the type system
  already enforces the absence on AgentConnector).
• src/index.ts re-exports bridge classes (LocalModelMcpConnector,
  DataStoreMcpConnector, SkillStoreMcpConnector) from root so the
  examples/custom-bootstrap.example.ts reference comment resolves
  without a sub-path import. Adopter-facing first-touch carry-over from
  e152851e.
• examples/onboarding-scaffold/bootstrap.ts + examples/custom-bootstrap.example.ts
  thread Registry into the LocalModelMcpConnector constructor +
  comment-document the registry-alias resolution.
• +8 tests in tests/v0.16.3-manifest-exposure.test.ts covering all
  three states across every substrate slot. 1434 tests passing
  (was 1426 at v0.16.2).

Counting

This is the 10th instance of the discipline-only-contracts class closed
since d0fb1375. v0.16 series total: 5 closures
(canonicalization / trust-boundary / parity / bridge-routing / manifest-exposure).

Process note — checkpoint-before-code

This ring validated the cold-pickup checkpoint discipline. Prior session
ran out of context with v0.16.3 scoped but uncoded; this session
reconstructed the plan from durable banks + filed it back for sign-off
before any code. Surfaced two contract details that would otherwise have
shipped wrong-shape: (a) AgentConnector has no manifest() per v0.9.6,
which invalidated the original "include 5 slots for symmetry" sign-off
assumption; (b) field-shape decision (separate manifest_unsupported
sentinel vs. unified manifest_error) needed an explicit choice for
dashboard consumers. Banked as a feedback memory; pattern continues to
earn its keep.