feat(harness): add host capability seams with inert defaults

[senamakel]

Summary

Adds harness::host — ten host-capability seams with inert default impls, plus the spec
that records why they exist and where their boundary sits.

A generic agent runtime needs a way for an embedding application to supply the capabilities
the runtime itself should not own: long-term recall, turn context assembly, security
decisions, budget admission, agent definitions, experience, learning, progress delivery,
tool-outcome classification, and model resolution. The crate already ships 18 harness
extension traits on exactly this pattern (ChatModel, Tool, ChatHistory, Store,
Summarizer, …); this is ten more of the same kind, not a new architecture.

Nothing in the crate calls these yet. They land empty on purpose, so a host can adopt them
incrementally without the runtime depending on the outcome. OpenHuman is the first consumer.

Spec: docs/spec/host-capabilities-spec.md. Module doc: docs/modules/harness/host.md.

Scope boundaries worth reviewing

These are the decisions that shape the surface — each is argued in the spec:

• BudgetGate is money and admission only. Context-window fitting is Summarizer +
  RunLimits, not a budget concern.
• ProgressSink is delivery, not transport. No receive side. An interactive loop that
  reads from a terminal or chat platform stays host surface.
• SecurityGate::filter_tools is advertisement, not enforcement. Per-call authorization
  is authorize_call, which is the only async method on the trait because it is the one
  site that genuinely awaits.
• AgentDefinition is deliberately minimal (id, description, system_prompt,
  model, tools, extras: Value) with Default derived. Richer host types map into it
  via extras rather than the crate freezing fields it cannot generalize.
• Configuration is not a seam. Every method takes &State, so a host parks a reloadable
  handle there and keeps read-at-call-time semantics with zero crate surface — better than a
  virtual getter, and it needs no new type.

Open question for reviewers

All ten are generic over State, matching ChatModel / Tool / Middleware. That may be
the wrong call, and I would like a maintainer's read before hosts adopt them.

Only 7 of the existing 18 extension traits carry State, and all 7 are execution traits.
The 11 that are genuinely capability traits — ChatHistory, Store, Summarizer,
EmbeddingModel, VectorStore, ResponseCache, WorkspaceIsolation, … — are bare
pub trait X: Send + Sync capturing their dependencies in Arc fields, and are the closer
analogue to these ten.

The alternative is: make the eight capability seams non-generic Arc<dyn Trait>, keeping
State only on ModelResolver (which must return ModelRegistry<State>) and
ContextComposer (closest to a middleware). Changing this later is a signature break for
every implementor, so it is cheaper to settle now.

API Or Behavior Changes

Additive only. No existing API changed, no behavior changed. Nothing in the crate
calls the new traits.

New public surface under harness::host, re-exported from the crate root:

Trait	Default impl
MemoryProvider	InMemoryMemoryProvider
ContextComposer	PassthroughContextComposer
SecurityGate	RootContainedSecurityGate
BudgetGate	UnmeteredBudgetGate
DefinitionRegistry	InMemoryDefinitionRegistry
ExperienceStore	NoopExperienceStore
LearningSink	NoopLearningSink
ProgressSink	NoopProgressSink
ToolOutcomeClassifier	NoopToolOutcomeClassifier
ModelResolver	StaticModelResolver

Version 2.1.0 → 2.1.1 (patch). Deliberate: the embedding host pins ^2.1, and a minor
bump would force a host manifest edit this change is scoped to avoid.

Most methods are sync, matching the call sites they are modelled on. Only
SecurityGate::authorize_call, ContextComposer::prepare_turn, and the MemoryProvider /
ExperienceStore / LearningSink I/O methods are async.

Tests

• cargo fmt --check
• cargo clippy --all-targets -- -D warnings
• cargo clippy --all-targets --all-features -- -D warnings
• cargo build --all-targets
• cargo build --all-targets --all-features
• cargo test
• cargo test --all-features

cargo test --all-features: 1330 lib tests, all integration targets, 45 doctests, 0 failures.
37 of those cover the new module's default impls.

Also added tests/host_seam_hygiene.rs, which scans src/harness/host/ for embedder-specific
vocabulary and fails with file, line, and reason. It runs under the normal test command rather
than as a CI lane someone has to remember, and its SCANNED_DIRS list is documented as needing
to grow as more code moves into the crate.

Documentation

• docs/spec/host-capabilities-spec.md — the catalogue, the rejected alternatives, and the
  publishability boundary. Filed in docs/spec/ per docs/spec/README.md, indexed from its
  Module 1 section.
• docs/modules/harness/host.md — per-trait scope table, including the three coupling
  concerns resolved by scoping rather than by adding an eleventh trait.
• Full rustdoc on every trait, method, and public field.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: f65381f1-25f8-4524-9cb8-6c42e11ec6f9

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}