feat(agent-spec): simplify resource references

[schickling]

Summary

• make reason mandatory on every Agent Spec Resource reference
• remove the generic relation field
• add optional inactive-reason so append-mostly Agent Specs can retain references that are no longer active for that agent

Motivation

Agent Specs are becoming the passive coordination index for agent workstreams. A local label is useful to the reading agent, while the URI carries the actual resource type and identity. The generic relation field duplicated the label/reason without establishing useful semantics.

Long-lived specs also need to retain historical references for traceability without implying that every referenced PR, worktree, or worklog remains current. inactive-reason expresses only that the reference is inactive in this Agent Spec; it does not alter the resource's lifecycle, authority, or resolution.

The resulting envelope stays intentionally small:

resource "coordination-pr" uri="github-pr://schickling/dotfiles/1887" reason="Owns the passive-coordination rollout."
resource "earlier-eval" uri="file:///.../eval.json" reason="Captured the baseline behavior." inactive-reason="Superseded by the live pilot."

Verification

• cargo test -p agent-spec - PASS (68 tests across unit/integration targets)
• cargo test --all-targets --all-features - Resource/parser changes PASS; one unrelated runtime test, incomplete_apply_marker_blocks_declarations_but_not_the_state_plane, is RED locally and was not modified by this PR
• cargo fmt --all -- --check - PASS

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 363805c80c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[schickling]

Rebased onto main (4629aeb — #309, the darwin capability-linkat fix). The branch previously sat on 9b1a3b4, which cannot build or test on macOS (linkat on /dev/fd/N → EPERM, see #308); consumers pinning this head were blocked from Darwin builds entirely.

No semantic changes: the rebase is clean (disjoint surfaces — agent-spec/crates vs event.rs/ding). Verified at a1931fb: Linux x86_64 cargo test -p st2 --lib 343 passed / 0 failed; macOS arm64 suite green (evidence to follow in dotfiles consumer verification).

Downstream note for schickling/dotfiles#1898: once this merges, one forward bump carries both the resource-reference envelope and the darwin fix.

[schickling]

macOS arm64 full-suite result at a1931fb: 336 passed; 1 failed — the single failure is the same pre-existing darwin parallel-load flake family in codex_app_server (subscribed_control_pump_delivers_a_typed_reference_to_the_real_fifo_head, Attempted≠Accepted); it passes in isolation (1 passed; 0 failed) and is untouched by this branch or #309. All ding/event/archive tests green.