Add observable Resource Profiles

[schickling-assistant]

Problem

Resource bindings can resolve passive filesystem carriers, but they cannot expose provider-owned current state or notify an agent when selected semantic state changes.

Goal

Add state-first, read-and-observe Resource Profiles with one canonical snapshot per binding, bounded profile-owned observation runtimes, semantic selectors, and level-triggered catch-up through the existing resync inbox and DING path.

Decisions

• Keep the Resource URI as identity and store attention configuration in a profile-validated JSON selector.
• Make one atomic current snapshot authoritative. Notifications are thin invalidations, not an event log.
• Keep provider authentication, observation, cursors, and reconciliation inside downstream runtimes.
• Fence publications with a runtime-incarnation owner claim and a binding registration token.
• Reuse the built-in resync stream with producer-side supersession.
• Bound selectors to 16 KiB, protocol lines to 2 MiB, decoded snapshots to 1 MiB, and health detail to 16 KiB.

Verification

• cargo check --workspace --all-targets --quiet
• Agent Spec descriptor tests: 7 passed
• Wasm descriptor/containment suite: 25 passed
• Runtime core tests: 11 passed
• Runtime supervisor tests: 5 passed
• Resource CLI integration tests: 8 passed
• Observable publication-to-inbox E2E: 1 passed
• nix flake check --no-build
• nix build .#checks.x86_64-linux.wasm-resolver-feature --no-link --print-build-logs passed on the rebased head.
• Strict VRS checks passed for 07-resource and 07-resource-profile.

Pre-flip deviation: nix build .#checks.x86_64-linux.st2 reaches 572 passing tests and then fails in the pre-existing release-mode Codex WebSocket handshake race with Interrupted system call (os error 4). A pristine origin/main derivation reproduced the same failure class in another Codex handshake test after 555 passing tests. The Resource Profile derivation and all affected focused tests pass.

Complexity

The change adds a resident runtime supervisor and durable publication state because provider-native observation crosses process and crash boundaries. It does not add another task plane, event plane, provider event log, or external dependency.

Concerns

Runtime health is currently available through the supervisor API. A dedicated status CLI projection remains follow-up dogfood work. Runtime exit is reconciled on the next resident catalog pass rather than by a separate restart timer.

Friction & bottlenecks

The repository has no devenv.nix, so the standard first-party check:quick task is unavailable. The repository-wide rustfmt check reports substantial pre-existing formatting drift outside this change. The release-mode Codex handshake suite also fails nondeterministically on pristine main as described above.

Follow-ups

• Use downstream GitHub-profile dogfood to settle the minimum low-noise operator health surface.
• Add frozen cross-version descriptor and runtime-protocol fixtures before independent third-party implementations.

References

Refs #376

Posted on behalf of @schickling

field	value
agent_identity	dev3.direct.omp.2cshu64q
session	dev3.2cshu64q
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.0.9
agent_runtime	OMP 18.0.9
tooling_profile	dotfiles@b607597