Remove legacy runtime autofix agent

[davejohnson]

What changed

• removed the standalone hourly Claude runtime-log autofix workflow and src/agent implementation
• removed its repo-local watch/error state, Anthropic SDK dependency, and npm entrypoint
• kept hv_logs and the live hv_errors list|summary provider-neutral observability paths
• removed hv_errors tracked|ignore, which only operated on the deleted local agent state
• made legacy environments.<env>.autofix intent fail with explicit migration guidance
• documented the boundary between runtime error visibility and generic GitHub workflow autofix

Why

Hypervibe now manages GitHub checks and autofix declaratively through github.actions. The older runtime agent mixed provider log credentials, model access, repository mutation, and state-file commits in one scheduled workflow. Keeping both systems would leave overlapping automation models and a broader secret surface.

This cleanup intentionally does not remove runtime error visibility. Users can still inspect recent production errors and per-service summaries through recorded provider connections. A scheduled runtime-error inbox or desktop alerting flow can be designed later as a separate desired-state capability.

User impact

• hv_errors action="list" and action="summary" continue to work.
• hv_errors action="tracked" and action="ignore" are removed.
• specs containing environments.<env>.autofix now receive actionable migration guidance.
• failed GitHub checks continue to use github.actions.<id>.kind="autofix" and review-gated draft PRs.

Validation

• 103 TypeScript test files / 808 tests
• TypeScript typecheck and production build
• prepublish sensitive-file/package safety check
• 35 macOS companion tests
• focused observability, schema, provider-log, and tool-contract tests
• package-lock JSON and dependency-tree validation
• GitHub Actions YAML parsing
• stale-reference and whitespace checks