GeezerVison

Orca publishes frequent releases with large Linux AppImage and Windows installer assets. I use the generic GitHub Release proxy to cache these downloads; fetching or prewarming new releases leaves older installers in storage, and their accumulated size can exhaust the available disk. storage.max_size provides a global LRU limit, but cannot express an inactivity limit for selected GitHub repositories.

This adds explicit opt-in retention for generic artifacts while preserving the behavior of existing storage.retention.default configurations.

Compatibility and configuration

The discussion in #306 deliberately excluded generic because release assets may disappear upstream. This PR preserves that behavior unless a user explicitly opts in. The global storage.retention.default never applies to generic, even when packages contains generic entries. Other ecosystems retain their existing behavior.

1. Preserve existing behavior: no opt-in. Leave storage.retention unset. Existing installations that already use only a global default also keep generic excluded:

storage:
  retention:
    default: "30d"  # Other supported ecosystems only; generic stays excluded.

2. Opt in selected packages only. This example enables retention for one GitHub repository. It does not enable retention for any other generic repository. default is optional and is omitted here to make the scope clear:

storage:
  retention:
    packages:
      "pkg:generic/owner%2Frepo": "30d"

Adding default: "30d" to this example would affect other supported ecosystems, but would still not opt in unlisted generic packages. A package rule covers that package's versions and assets; each cached asset ages individually based on its last access, with recently fetched files protected as before.

3. Opt in all generic packages. Set the generic ecosystem rule explicitly:

storage:
  retention:
    ecosystems:
      generic: "30d"

4. Override or exclude a package after opting in the ecosystem. This keeps one repository outside age-based eviction and gives another a longer retention:

storage:
  retention:
    ecosystems:
      generic: "30d"
    packages:
      "pkg:generic/owner%2Fkeep": "0"
      "pkg:generic/owner%2Flong-lived": "90d"

For generic, precedence is matching package rule → explicit generic ecosystem rule → no age-based eviction. There is no fallback to the global default.

Explicit rules also apply to existing cached assets, so the initial sweep can evict older unused files. Users opting in accept that an evicted upstream asset may no longer be available for refetch. storage.max_size remains independent: exclusion from age-based eviction, including "0", does not prevent capacity-based LRU eviction.

Implementation

• Register generic with its existing stored package PURL form, such as pkg:generic/jqlang%2Fjq.
• Resolve explicit package and ecosystem rules before returning zero for generic instead of falling back to the global default.
• Keep generic in the startup log's default_not_applied list.
• Reuse the existing sweeper, deletion queue, and grace period, without new configuration fields or schema changes. Metadata/API responses are not retention candidates.
• Document the opt-in behavior and upstream availability tradeoff.

Validation

Tests cover legacy global-default preservation, explicit ecosystem and package opt-in, isolation from other generic repositories, package-level disabling, actual Release package PURL storage, recent-access protection, eviction, and refetch through the generic HTTP route. The integration upstream is an httptest.Server, not live GitHub.

Passed on Linux with Go 1.26.7, against a source snapshot containing only this PR's changes:

go test -p=1 ./... -timeout 5m
go vet -p=1 ./...
go build -p=1 -buildvcs=false ./...
go test -p=1 -race ./internal/config ./internal/retention ./internal/handler ./internal/server -run 'Generic|Retention|NotCovered|RulesFor|RulesMinPositive' -count=1 -timeout 3m

Changed Go files pass gofmt, and git diff --check passes. -buildvcs=false was used because the validation snapshot has no .git directory. PostgreSQL integration tests were skipped (PROXY_DATABASE_URL was unset); local tests used SQLite. Full race and lint checks across the CI operating-system matrix are left to CI.

Related to #306 and #428.