Do not replace registry errors with "<port> does not exist"

[Bugale Bugalit (bugale)]

Fixes #2094.

VersionedPortfileProviderImpl::entry() threw away the error returned by get_port_entry and substituted {package_name} does not exist, so any registry-level failure was reported as the port not existing.

get_port_entry returns ExpectedL<std::unique_ptr<RegistryEntry>> with three outcomes:

1. error — something went wrong talking to the registry
2. value, null pointer — the registry genuinely has no such port
3. value, non-null — found

Case 2 is already handled downstream in load_control_file:

if (!ent->get())
{
    return msg::format_error(msgPortDoesNotExist, msg::package_name = version_spec.port_name);
}

so the else branch touched here is reached only for case 1. It also used msg::format rather than msg::format_error, so the line wasn't even prefixed with error:.

Before / after

This hid #2066 from us entirely. A concurrent git init race in the registries cache made get_port_entry fail, and the entire user-visible output was:

Fetching registry information from https://github.com/microsoft/vcpkg (HEAD)...
apr does not exist
apr-util does not exist
capstone does not exist
crc32c does not exist
expat does not exist
...

No error: prefix, no git output — just every port from the default registry declared non-existent, which points suspicion at the manifest or the baseline instead of at the registry fetch that actually failed. Finding the real message required a GIT_TRACE2_EVENT capture of a live failure:

could not lock config file .../registries/git/.git/config: File exists
could not set 'core.repositoryformatversion' to '0'

With this change that message reaches the user directly.

Worth noting: #2066 hit the same underlying failure but reached it through get_baseline_version, which propagates errors correctly — so that reporter saw the actual git error. The difference was entirely this one else branch.

The change

entry_it = m_entry_cache.emplace(name.to_string(), std::move(entry).error()).first;

ExpectedT::error() && already exists and the cache value type is constructible from LocalizedString, so this is a drop-in replacement for the discarded msg::format(...). msgPortDoesNotExist remains in use for the genuine cases at both other call sites in this file.

[Copilot]

Pull request overview

This PR fixes misleading error reporting in the versioned registry port lookup path by preserving and surfacing the actual registry error returned from RegistryImplementation::get_port_entry(), instead of replacing it with a generic "<port> does not exist" message.

Changes:

• In VersionedPortfileProviderImpl::entry(), propagate the ExpectedL error (std::move(entry).error()) into the cache rather than formatting msgPortDoesNotExist.
• As a result, registry-level failures (git/network/filesystem/locking) now reach users directly via the existing ExpectedL error propagation path.

[Billy O'Neal (BillyONeal)]

Is this whole thing now just:

entry_it = m_entry_cache.emplace(name.to_string(), reg->get_port_entry(name)).first;

I believe what this was intending to do was preserve the error cases and map the "nullptr unique_ptr" case to the "port does not exist" error message. So I think it's good to not drop the "child" error but I think this is incorrectly dropping a legitimate error too.

[Billy O'Neal (BillyONeal)]

I missed that you already discussed this in the description but the "is this whole thing now just" comment remains

[Bugale Bugalit (bugale)]

Billy O'Neal (@BillyONeal)
Yeah, you're right. Once the error is kept, both branches store the same thing, so the if/else does nothing. Pushed your version.
The nullptr case still gets the "does not exist" message, just one level up: entry() is only called from load_control_file, and that checks !ent->get() right after unwrapping. So a genuinely missing port reads the same as before, and a git or network failure now says what actually broke.

Oh I see my comment is addressed in the description

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.