feat(examples): add deterministic oracle sweep

[christso]

Summary

Adds a deterministic oracle sweep for agent/LLM-backed example evals so maintainers can prove those examples still load, execute, and write artifacts without provider credentials or live LLM calls.

This is a contract oracle for example execution, not a set of captured live-model golden transcripts. Examples that already use oracle-capable targets are classified as covered and are not replay-overridden.

Oracle Definition and Coverage

In this PR, an example eval's oracle is defined by its target kind:

• Oracle-capable targets are their own oracle. These include mock, cli, replay, transcript, and passive copilot-log providers. The sweep inventories them as oracle_target and does not replace their target output.
• Agent or LLM-backed targets require an oracle fixture. These include concrete model providers, interactive agent providers, and unresolved env-delegated targets such as root llm -> ${{ LLM_TARGET }} or default -> ${{ AGENT_TARGET }}. The sweep generates replay fixture rows for these evals and runs them with --target example_oracle plus --grader-target example_oracle_grader.
• Generated replay outputs come from existing reference_answer / expected_output when available, otherwise from deterministic assertion-derived content. The goal is contract coverage for parsing, execution, grading plumbing, and artifact writing without provider keys, not semantic live-model quality scoring.

Current coverage from bun run examples:oracle -- --inventory:

• 113 eval files discovered under examples/
• 85 requires_oracle_fixture: run through generated replay fixtures
• 24 oracle_target: already covered by oracle-capable local/static/replay targets
• 0 needs_fixture_added
• 4 excluded with explicit reasons below

What changed

• Adds bun run examples:oracle as the standard workflow for running non-oracle example evals against a generated replay target and oracle CLI grader target.
• Adds bun run examples:oracle:gate as the modular gate wrapper: it runs validate:examples, then forwards arguments to examples:oracle.
• Adds examples/oracle-fixtures.yaml as the maintained oracle inventory control file with explicit exclusions.
• Classifies example eval targets as requires_oracle_fixture, oracle_target, or excluded.
• Generates replay fixtures from reference answers, expected assistant outputs, or assertion-derived deterministic content at runtime under .agentv/tmp/ only for evals that otherwise require an agent or LLM target.
• Adds a publish-next GitHub Actions gate that runs bun run build once, then bun run examples:oracle:gate, before npm publish. The next publish path then calls bun scripts/publish.ts next directly to avoid a second build.
• Removes stale examples/features/workspace-setup-script/evals/dataset-vscode.eval.yaml because VSCode target support was removed.
• Updates the cross-repo TypeScript runner to prefer Bun for .ts scripts, avoiding an unnecessary tsx dependency during offline oracle runs.
• Documents the oracle sweep in examples/README.md.

Exclusions

• examples/features/docker-workspace/evals/docker-example.EVAL.yaml: requires Docker workspace runtime; replaying target output does not remove Docker setup dependency.
• examples/features/copilot-log-eval/evals/skill-trigger.EVAL.yaml: requires local Copilot session/workspace artifacts and .github/skills/agentv-bench/SKILL.md during before_all setup.
• examples/features/prompt-template-sdk/evals/dataset.eval.yaml: loader returns zero runnable tests because the eval references missing ../prompts/custom-grader.ts command file.
• examples/showcase/bug-fix-benchmark/evals/bug-fixes.eval.yaml: invalid eval metadata today; experiment targets[0].use_target is empty.

Verification

• bun install
• bun run build
• bun run examples:oracle -- --eval examples/features/assert/evals/dataset.eval.yaml --eval examples/features/default-graders/evals/dataset.eval.yaml --eval examples/showcase/grader-conformance/EVAL.yaml --output-dir .agentv/tmp/example-oracle-smoke
• bun run examples:oracle -- --output-dir .agentv/tmp/example-oracle-full-2 passed before target classification narrowing: 110 runnable, 0 failures, 4 excluded
• bun run examples:oracle -- --output-dir .agentv/tmp/example-oracle-agent-targets passed after target classification narrowing: 86 oracle-required, 0 failures, 24 oracle-capable targets, 4 excluded
• bun run build && bun run validate:examples && bun run examples:oracle -- --output-dir .agentv/tmp/example-oracle-publish-next-check passed after removing stale VSCode eval: 85 oracle-required, 0 failures, 24 oracle targets, 4 excluded
• bun run examples:oracle:gate -- --inventory
• bun run validate:examples
• bun run examples:oracle -- --inventory
• bun run lint
• bun run typecheck
• actionlint not installed locally; workflow syntax was reviewed by inspection

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	30f8d92
Status:	✅  Deploy successful!
Preview URL:	https://d70bbf7f.agentv.pages.dev
Branch Preview URL:	https://feat-example-oracle-fixtures.agentv.pages.dev

View logs

[christso]

Superseded by Bead av-rjz6. Scope is reduced to oracle/replay only for expensive agent targets while keeping declared graders unchanged; this draft used mocked grader plumbing and should not be merged.