eval: select and prove native Agent Spec resources

[schickling-assistant]

Problem

Agent Spec Resource bindings need both behavioral selection evidence and native runtime acceptance before they can replace Assignment or Focus wrappers. The prior tournament compared the three treatments, but it did not prove the actual catalog envelope implemented by merged st2 #86.

Goal

Provide the executable Agent Spec companion to st2 #86: select direct named Resources, document the exact portable envelope in AGENT-SPEC.md, and prove it through free, maintained runtime evidence.

Authority boundary

This restack preserves the canonical-contract wording merged by PR #65 exactly: evals owns AGENT-SPEC.md and its maintained acceptance cells as the contract and proof surface; st2 is the current implementation, not the contract owner; the implementation pin identifies only the version currently proved; contract and proof-cell updates precede conformance claims.

Relative to previously reviewed PR #45 head 82f04778a0624963b95145de89c96e472ff73063, the semantic tree change is exactly PR #65's three documentation edits: AGENT-SPEC.md, README.md, and docs/vrs/spec.md (22 additions, 12 deletions). PR #45's product/eval behavior is unchanged.

Decisions

• Direct named Resources remain the selected treatment; Focus and Assignment remain matched controls.
• The portable envelope is only resource "<name>" _tag="<opaque-type>" uri="<absolute-uri>".
• The binding name is an agent-local semantic role, the URI is exact identity, and declaration order is not semantic.
• Unsupported access, readiness, required/optional, resolution, and lifecycle properties fail closed rather than appearing enforced.
• Folder-eval Resource projection is not inferred from the catalog envelope.
• Resource-only declaration edits update inspection while a healthy live task is adopted without replacement.

Changes

• Add a model-free native Resource-binding cell with KDL/TOML/JSON parity and negative fixtures.
• Keep the direct-Resource, Focus, and Assignment cold/hot/handoff tournament as matched evidence.
• Materialize treatment-specific worker AGENTS.md loaders for the three cold-start cells.
• Pin the accepted published st2 runner and its immutable release provenance.

Published runner

• source: 0fed14bb5653b67e1d64f1199e240c4c5c612bf7
• release: v0.2.0+0fed14b
• terminal-green publication: release-portable run 30550227417
• archive SHA-256: d14404ae678bbe3f2a5ad8580cde1e4b8f6009067c46555f392c6e0957b8a2da
• extracted binary SHA-256: d61d12b2b1189a391c196ca28f8f4ba69072d14fcbad2571fc29db1f250f4eed
• downloaded SHA256SUMS SHA-256: 50cfd8722e58d1c74fdc543f3e3bb3bac768decd04575fde2360ea838ec5e9d3

Verification

Restacked on current main commit e9b53e79b05b1c0e1d7eea02db2eaba47376fe05.

• native agent-spec-resource-bindings: 5/5 PASS, including strict failures, deterministic inspection, Resource-only adoption, and zero-PTY/process cleanup
• all nine matched tournament oracle self-tests PASS
• 220 shell files parse
• 84 explicit model launches remain cost-controlled; no model was launched
• 83 model seats remain event-first and harness-backed
• 107 KDL files parse
• 57 maintained cells and 11 nested catalogs load
• 17 derived model-free cells have explicit exclusions
• all fixture-reset, VRS, Weird-Git, closed-set mutation, overnight-policy, no-PII history, generated-catalog, and diff gates pass
• terminal result: PASS: complete model-free corpus preflight

No provider/model call or deployment was made during this restack.

Scope boundary

Resource resolution, access grants, required/optional status, readiness, lifecycle, mutation, and rendering remain outside the generic envelope. The URI's presence grants no authority.

Refs #41.

[myobie]

Exact review green at 2c259bb. Tree comparison against reviewed head 82f0477 changes only AGENT-SPEC.md, README.md, and docs/vrs/spec.md; those bytes are the already merged PR65 authority boundary. Product/eval behavior is unchanged, the complete model-free preflight is green, and no provider call ran.