feat: ship a Claude Code plugin with the user skills

[aaaaahaaaaa]

Summary

A Claude Code plugin aimed at users of the framework (people with interloper-core in their own project), distinct from the contributor skills in .claude/skills/.

• .claude-plugin/marketplace.json: the repository is installable as a marketplace (/plugin marketplace add digitl-cloud/interloper, then /plugin install interloper@interloper).
• plugins/interloper/: the plugin manifest (version pinned to the framework release) and ten skills:
  • interloper-source: writing a source with a connection, schemas on every asset, partitioning, dependencies, verification on one partition, typed read-back, registration.
  • interloper-run: run() versus materialize(), the DAG, the sync bridge, the CLI recipe and its spec file.
  • interloper-connection: credentials, OAuth through a provider, check(), fetch-field pickers, definition() attributes.
  • interloper-schema: what AUTO, STRICT and RECONCILE do, normalizer options, value-level transforms, error signatures.
  • interloper-destination: PartitionedDestination and DatabaseDestination hooks, window splitting, REPLACE semantics, reconcile on read.
  • interloper-manifest: the job spec shape, ${VAR} rules, select versus the assets whitelist, cross-source wiring by id.
  • interloper-backfill: one run per partition, windowed runs, lookback windows, per-partition progress and outcomes.
  • interloper-triage: reading an event stream, placing the failure in the operation lifecycle, error signature to fix table, when a retry helps.
  • interloper-deploy: interloper.yaml and INTERLOPER_* precedence, encryption, sign-in, catalog, provider apps, telemetry, images and chart.
  • interloper-upgrade: version discovery, changelog reading, the rename table from FileIO and OAuthConnectionBase to today's names.
• docs/reference/claude-code.md: install instructions and the skill list, linked from the Reference nav.

Method

Every skill follows the same discipline: a subagent played a framework user without the skill (public docs and introspection only) on a realistic task, its report gave the failure list, the skill was written against that list, and the scenario was re-run with the skill and refined from the residual gaps. Ten scenarios in total (source authoring, run and backfill, OAuth connection with picker, vendor payload typing, JSONL and SQLite destinations, a two-source job spec, daily plus monthly backfill, three failed-run event logs, a platform deployment, a legacy script upgrade).

Docs corrections the baseline runs exposed

• Specs: the assets override map is a whitelist; assets absent from it do not exist after reconstruction.
• Dependencies: cross-source requires is a contract check, not wiring; the edge is wired by id in Python or in the spec.
• Settings and Running the app: interloper.yaml is not interpolated, a field present in the YAML wins over its environment variable and the environment fills the rest.
• Destinations: the bare il.Destination example is not partition-safe; partitioned assets belong on PartitionedDestination.

Follow-ups (not in this PR)

• Resolve cross-source requires in DAG.__init__ instead of requiring manual wiring (found while testing; a task chip was raised).
• Bumping plugin.json / marketplace.json versions from semantic-release alongside the pyproject files.

By Digitl