Hosted/vendored NuGet mapping isn't exclusive when nuget.config already maps the exact package id to another source, so restore races the Socket feed against nuget.org (NU1403 with a lock, silently unpatched without one)

[mikolalysenko]

[agent] Found by the scheduled NuGet / dotnet bug-hunt routine (ledger #320).

Summary

When the project's nuget.config already has a <packageSourceMapping> that names the patched package's exact id under another source (Visual Studio's package-source-mapping UI writes exactly this, e.g. <package pattern="Newtonsoft.Json" /> under nuget.org), scan --mode hosted (and --mode vendored) just adds a second exact-id mapping under socket-patch-<uuid>. Two sources now have the same most-specific pattern, so NuGet treats both as eligible and takes the package from whichever answers first.

The rewriter assumes "most specific pattern wins" (crates/socket-patch-core/src/patch/redirect/mod.rs:6696). That doesn't hold for a tie, and nothing detects the tie, warns about it, or removes the user's entry.

Impact

• With packages.lock.json: if nuget.org answers before the Socket feed, a fresh dotnet restore --locked-mode downloads the upstream nupkg and fails with NU1403: Package content hash validation failed. Whether CI is green or red then depends on feed latency.
• Without a lock: the same restore exits 0 and installs the unpatched upstream bytes. scan reported success and printed no warning.

Repro

This uses the repo's own real-dotnet harness (crates/socket-patch-cli/tests/e2e_nuget_dotnet_build.rs: its Backend wiremock stand-in for the API and the hosted feed, a real nuget.org fixture restore, and fresh_checkout). After restore_fixture, overwrite the fixture's config with the one below and restore it once (--locked-mode, which passes):

<configuration>
  <packageSources>
    <clear />
    <add key="nuget.org" value="https://api.nuget.org/v3/index.json" />
  </packageSources>
  <packageSourceMapping>
    <packageSource key="nuget.org">
      <package pattern="*" />
      <package pattern="Newtonsoft.Json" />
    </packageSource>
  </packageSourceMapping>
</configuration>

Then run:

1. socket-patch scan --mode hosted --json --yes --api-url <stub> --org test-org --api-token x --patch-server-url <stub> exits 0, warnings: [].
2. Make a fresh checkout (csproj, nuget.config, lock), then run dotnet restore --locked-mode with a cold NUGET_PACKAGES and a cold http cache.
3. Make the stand-in's flat-container index.json answer 3 s late (ResponseTemplate::set_delay), standing in for a Socket feed that is slower than nuget.org's CDN.

Config after step 1:

  <packageSourceMapping>
    <packageSource key="socket-patch-4e4e4e4e-…">
      <package pattern="Newtonsoft.Json" />
    </packageSource>
    <packageSource key="nuget.org">
      <package pattern="*" />
      <package pattern="Newtonsoft.Json" />      <!-- still there: a tie -->
    </packageSource>
  </packageSourceMapping>

Results (2/2 runs, 2 cold restores each):

shape	Socket feed delay	result
user exact-id mapping, locked restore	3000 ms	error NU1403, 0 requests for the Socket nupkg, cache holds the pristine LICENSE.md
user exact-id mapping, lockless restore	3000 ms	exit 0, pristine bytes installed
user exact-id mapping, locked restore	0 ms / 500 ms	pass (the local stand-in wins the race)
control: default nuget.config (no user mapping), locked restore	3000 ms	pass (the Socket source is the only eligible one)

The control shows the failure comes from the tie, not from the delay. Vendored mode writes the same tie (socket-patch-<uuid> → .socket/vendor/nuget/<uuid> beside the user's nuget.org → Newtonsoft.Json). Its local folder feed usually wins the race, so I didn't get a failing vendored restore. The config is equally non-exclusive, though.

Expected vs actual

• Expected: CLI_CONTRACT.md (hosted/vendored table, nuget row) says the rewrite adds "source socket-patch-<uuid> + its exclusive exact-id <packageSourceMapping>". docs/ecosystems.md (NuGet row) says hosted pins the id via source mapping plus the lock contentHash. The patched id should resolve only from the Socket source. So the rewriter should drop or override the user's same-id pattern on other sources, or at least refuse or warn.
• Actual: the user's exact-id entry stays, the mapping isn't exclusive, restore is nondeterministic, and the lockless case silently installs the unpatched package.

Matrix

OS	SDK	hosted	vendored
Linux	8.0.131	fail (locked: NU1403; lockless: silently unpatched)	config tie reproduced; restore passed (local feed wins)
macOS / Windows, SDK 6/9/10		untested (SDK downloads from dot.net are blocked in this sandbox). The race is NuGet client behaviour, so I expect the same result	untested

Main 2463257 (v5 consolidation, #277).

Suspect code

• crates/socket-patch-core/src/patch/redirect/mod.rs:4132-4140: when a mapping already exists (!creating_mapping), it appends only the Socket <packageSource> and never looks for the same id under other sources.
• crates/socket-patch-core/src/vendor/nuget_feed.rs:896 / :915: the same append for vendored.

Related, but not the same: #354 (the * fan-out cuts off inherited sources) and #352 (a warm cache shadows the patch).

[mikolalysenko]

[agent] Triaged as priority:p3 (NuGet). The packageSourceMapping rewriter adds a Socket exact-id pattern without removing or detecting a user exact-id pattern for the same id under another source, producing a tie. No duplicate or existing fix PR found.

---

Generated by Claude Code