Cache network-backed resolutions and add refresh

[t-kalinowski]

Summary

Cache dependency graphs that pak resolves to network-backed package sources, so a second invocation of the same script can reuse the resolved library without launching R or performing another network operation.

Previously, the cache policy classified the user’s input ref before resolution. That excluded remote refs such as r-xla/anvl and could not account for transitive or mixed local/network graphs. This change lets pak resolve the graph once, then bases the cache policy on pak’s resolved source locations.

User-facing changes

• Reuse warm resolutions for any graph containing a network-backed source, including registry snapshots, GitHub, GitLab, Bitbucket, git, URL, SCP-style SSH, transitive remote dependencies, and mixed local/network graphs.
• Add --refresh to run, render, preview, tool run, rx, and tool install.
• Support nonempty IR_REFRESH as the environment-level refresh control.
• Treat IR_LATEST_RESOLUTION_MAX_AGE_SECONDS=0 as an unconditional miss for time-bounded markers, including markers written during the current second.
• Continue resolving local-only graphs and refs with pak query modifiers on every invocation.
• Preserve the previous warm result when a refresh fails.

Internal changes

• Remove custom Rust and R parsing of user package refs for cache policy decisions.
• Use pak’s resolved sources, mirror, and params fields to identify network-backed graphs and explicit query modifiers.
• Apply one expiring-marker policy to every network-backed R graph. exclude-newer remains part of the cache key but no longer makes a network resolution permanently current.
• Create marker timestamps only after successful materialization.
• Store each tool provider’s timestamp, library path, and package name in an independent marker so providers sharing a dependency set remain coherent across refreshes.
• Bypass both R and Python resolution markers during refresh, including reads after acquiring the resolver lock.
• Use pak’s direct-package result, with a focused pak solve when needed, to identify the primary package for tool metadata.
• Update command help, cache documentation, and public CLI regression coverage.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 66742c4a79

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".