feat!: one Relation primitive for every component link

[aaaaahaaaaa]

Summary

This PR replaces every link between components (destinations, connections, upstream assets, job
targets, hook watches) with one primitive: Relation. Today those links are five separate
mechanisms with five separate names (relation_types, RelationDefinition, resource_types,
ResourceRef, depends_on, Dependency, RelationSlot), each with its own selection rule,
fallback behaviour and serialisation quirks. Relation collapses all of them into one class,
declared the same way on every kind (an annotation, a typed or bare il.Relation attribute, or the
decorator's relations=), selected the same way (an exact key, source.asset, *.asset, or a list
of keys), and read the same way (bound value if wired, else a fallback resolved at read time, never
at bind time).

The point of unifying this is the north star the whole redesign serves: a campaign matcher asset
that fans in every campaigns asset an organisation has configured, with the set of upstreams
decided by wiring at the instance level, not hardcoded into the class. Assets get one addition on
top of the shared model: data() receives il.Upstream (or list[il.Upstream]) for asset
relations instead of the bound component, since an asset's data is what downstream assets actually
need, not the component itself. This PR is core (interloper-core, interloper-assets
connectors), plus the interloper-app and interloper-db files the earlier dependency ->
upstream rename touched; the platform, matcher and app follow as stacked PRs on top of it.

Decisions

Topic	Decision
Primitive	One Relation class declares every link, on every kind. It replaces relation_types, RelationDefinition, resource_types, ResourceRef, depends_on, Dependency and RelationSlot.
Selection	Always by key: an exact key, source.asset, *.asset, or a list of keys. A class in a declaration is shorthand for its key.
Bare key	A bare asset key selects a sibling of the same source instance.
Wiring	Always an instance on the component. Ids exist only in specs and in rows.
Fallback	Fallbacks are resolved when a relation is read, never bound: an explicit default= first, otherwise the declared class when it is a Resource (settings may come from the environment, so it is constructed at read time and a validation error surfaces then, as today) or when all its fields are defaulted. An unbound non-optional relation with no fallback is a build error.
Optional	Means the wiring may be absent. It says nothing about data: a bound upstream whose partition holds nothing yields data=None, for every relation.
Removal	on_delete is declared per relation (block or detach). Unbinding is derived: refused when it would empty a non-optional relation, detach otherwise. on_unbind is dropped.
Declaration forms	Annotation with a Component class (connection: Conn), a typed il.Relation attribute (destinations: list[Destination] = il.Relation("destination", many=True, optional=True), the annotation is for the type checker), a bare il.Relation attribute, or @il.asset(relations={...}). Explicit beats annotation; a subclass replaces its parent's declaration of the same name. Relation is its own descriptor (class access gives the Relation, instance access the bound value); a TYPE_CHECKING-only __new__ returning Any makes the typed form check under ty.
Injection	data() receives the bound instance for resource relations and il.Upstream (single) or list[il.Upstream] (many) for asset relations.
Serialisation	A manifest is one component, nested, and is exactly to_spec() output. A target that has a parent is emitted under its parent and referenced as {ref: id} everywhere else. A target without a parent is inline at the first relation that reaches it and referenced afterwards. No flag, no kind knowledge.
Containment	Stays parent_id (assets under sources). parent moves onto Component so the serialisation rule is generic. To revisit once this lands (containment as an owned relation).
Rows	component_relations(src_id, name, dst_id, ...), primary key (src_id, name, dst_id). type and slot collapse into name.
DAG	Edges from bound asset relations. A bound upstream not among the DAG's workloads is included as a read-only node.
Migration	Revision 017 is rewritten in place (it never shipped); no 018.
Compatibility	None. All consumers ship together; the manifests repo updates in lockstep.

Execution rulings, finalized during implementation

A few points the design left open were settled while building this PR:

• Declaration forms. Relation is its own descriptor and also carries a TYPE_CHECKING-only
  __new__ returning Any, so the typed attribute form (x: list[Destination] = Relation(...))
  type-checks under ty as if x were assigned the target type directly. Relation.__set__ owns
  rebinding, and Component.__setattr__ routes relation-name assignment to it, since pydantic's own
  __setattr__ only special-cases properties.
• Fallback. A Resource (BaseSettings) target is always self-filling. Its fallback constructs
  at read time and a ValidationError surfaces then if a required, env-backed field is missing; an
  unbound connection is no longer a build-time error, matching today's behaviour.
• Construction-time validation. validate_relations at construction time skips an unbound,
  non-optional asset relation whose key is not source-local (cross-source or wildcard keys can only
  be filled by the DAG, not by the source building its own assets). The DAG's own
  _check_relations still validates every live node once resolution has run.
• Read-only inclusion. A bound upstream not among the DAG's workloads joins as a single,
  non-transitive pass: it is read, not run, so its own upstreams are not pulled in behind it.
• Removal. Unbinding is derived from cardinality, not a separate on_unbind hook: refused when
  it would empty a non-optional relation, a plain detach otherwise.

What changes for connector authors

Before:

@il.source(resources={"connection": FacebookAdsConnection})
class FacebookAds(il.Source):
    account_id: str = il.InputField(discriminator=True)

    @il.asset(depends_on={"campaigns": il.Dependency(key="campaigns")})
    def campaigns_stats(self, context, connection, campaigns) -> list[dict]: ...

After:

@il.source
class FacebookAds(il.Source):
    connection: FacebookAdsConnection
    account_id: str = il.InputField(discriminator=True)

    def campaigns_stats(self, context: il.ExecutionContext, campaigns: il.Upstream) -> list[dict]: ...

The connection moves from a decorator kwarg to a plain annotation on the source; every asset that
needs it gets it trickled down automatically. The upstream moves from depends_on={...} with an
explicit il.Dependency to a bare il.Upstream-annotated parameter: the parameter's own name is the
bare key, so campaigns on campaigns_stats resolves to the sibling campaigns asset with no
further declaration. Cross-source or wildcard upstreams, and any relation that needs a non-default
key, still go through relations={...} on the decorator, unchanged in spirit from depends_on=.

All 24 connectors in interloper-assets have been moved onto this form.

Stack

This PR is phase 1 (core: interloper-core, interloper-assets connectors). Phase 2
(platform: interloper-db, interloper-api, interloper-toolkit, interloper-agent,
interloper-scheduler) is stacked on top of this branch as its own PR, followed by phase 3
(matcher) and phase 4 (app), each stacked on the one before.

It is not, however, strictly core-only: the branch also carries the earlier dependency ->
upstream rename commits, which touch interloper-app's catalog and component types and
interloper-db (migration 017 in the old type/slot shape, plus the relation store and its
tests). Phases 2 and 4 rewrite both of those, so what lands here is an intermediate state of
files those phases own, not their final shape.

Until phase 2 merges, the platform packages' test suites are red against this core: they still
reference the retired names (RelationDefinition, Dependency, resource_types, .upstreams,
.slots) that phase 1 removes. Alone, this branch leaves 37 platform tests failing
(interloper-api 4, interloper-scheduler 33) and 3 interloper-db test modules
uncollectable. This is expected and owned by phase 2, not a regression to chase down here.
Because of that, this PR must merge together with, or immediately before, phase 2 so main
is never left on a red platform for longer than the merge itself takes.

Verification

Counts below are after the review fix wave (one write path for bindings, id-based trickle
detection on copies, forbidden unknown Relation kwargs, refused uncollected component
annotations, fetch-provider validation on kind-declared relations).

• uv run --frozen ruff check: all checks passed. The pre-existing import-order issue in
  interloper-db/src/interloper_db/store/relations.py introduced by this branch's commit
  77598513 (an identity rename) has been fixed with ruff check --fix.
• uv run --frozen ty check: 18 diagnostics repo-wide, all in phase-2-owned packages and all
  expected (old relation names not yet ported): interloper-db 15, interloper-slack 2,
  interloper-scheduler 1. Zero in interloper-core, interloper-assets, interloper-google-cloud,
  interloper-pandas, interloper-docker, interloper-k8s.
• uv run --frozen pytest packages/interloper-core packages/interloper-assets packages/interloper-google-cloud packages/interloper-pandas packages/interloper-docker packages/interloper-k8s -q:
  1734 passed.
• Phase-2 packages, collected/run against this core for the record (expected to fail until phase 2
  lands): 37 failures and 3 uncollectable modules in total.
  • interloper-db: 3 collection errors (AttributeError on retired names in
    test_components.py, test_hydration.py, test_relations.py).
  • interloper-api: 4 failed, 380 passed.
  • interloper-toolkit: 10 passed (no coverage of the retired surface yet).
  • interloper-agent: 11 passed (no coverage of the retired surface yet).
  • interloper-scheduler: 33 failed, 78 passed.
• Retired-name sweep (resource_types, relation_types, RelationDefinition, Dependency,
  ResourceRef, depends_on, RelationSlot, AssetIdentity, and related old names) over
  interloper-core, interloper-assets, interloper-google-cloud, interloper-pandas,
  interloper-docker, interloper-k8s, docs, plugins, examples yields 12 hits across four
  categories, all deliberate:
  • 1 regression test (tests/component/test_base.py:845): asserts the old names are gone from the public API.
  • 4 docstring and comment hits (tests/resource/test_fields.py:98-99,109 and resource/fields.py:365): explain the current design by contrasting it with the retired one.
  • 5 upgrade skill migration table hits (plugins/interloper/skills/interloper-upgrade/SKILL.md:65,73-75,81): legitimate reference in the "Old" column of the 0.7x migration table.
  • 2 Docker compose syntax hits (examples/telemetry/docker-compose.yml:16,61): false positive on Docker's own depends_on: key.
• Em-dash sweep over interloper-core/interloper-assets source and the guide/extending/reference/ui
  docs: all em-dashes in files this branch touched were rewritten; the remaining hits are all in
  files this branch never touched (pre-existing house style elsewhere in the codebase, left alone).

Follow-ups

• Multi-dot declared keys resolve via rpartition; the semantics are unstated and untested.
• Unbinding an already-empty, non-optional relation is a silent no-op.
• Relation.__new__ returning Any means ty no longer checks Relation(...) call sites, and a
  non-optional relation annotated with the bare target type reads as always-present to ty even
  though it can be unbound at runtime.
• tests/source/test_base.py's Finance.Revenue docstring rationale for optional=True is stale;
  tests/dag mixes Matcher and FakeMatcher fixture naming inconsistently.
• A job-trickled destination serialises inline under its first target rather than under the job
  itself.

By Digitl

[codecov]

Codecov Report

❌ Patch coverage is 92.34973% with 14 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...per-scheduler/src/interloper_scheduler/executor.py	11.11%	8 Missing ⚠️
...kages/interloper-core/src/interloper/asset/base.py	96.73%	3 Missing ⚠️
...terloper-toolkit/src/interloper_toolkit/lineage.py	25.00%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!