Pipenv recognizes hosted PyPI patch URLs with two private grammars that disagree with the shared one #563

[agent] Filed by the scheduled architecture audit routine (ecosystems and formats). Register: #560 (comment).

Kind: bug. Source: review Part 5.4 ("Hosted-URL recognition is duplicated"), register E04. The hosted-side half is a new finding, register E49.

Problem

Four functions decide whether a URL in a Python lock is a Socket-hosted PyPI patch reference:

Reproduced on d63ae5f (unit tests, run twice):

• Hosted Pipenv cannot rotate its own pin on a path-prefixed patch server. With artifactUrl = https://patches.internal.example/socket/patch/pypi/urllib3/1.26.18/token/patch-one/urllib3-1.26.18-py3-none-any.whl, owned_url(artifact_url, dep) is false for the URL it just wrote. The second plan() (grant rotated) returns Conflict("Pipenv source for urllib3 already exists") (redirect/pipenv.rs#L355-L360).`` On the same URL hosted_artifact_url returns `uuid_level = Some("patch-one")`, so VEX and inventory treat the entry as Socket's while hosted treats it as a user source. The same origin without a path prefix (`:8443`) rotates fine.
• Vendored Pipenv calls a foreign host's URL Socket-hosted. is_socket_hosted_reference("https://evil.example/patch/pypi/requests/2.28.1/<grant>/<uuid>/requests-2.28.1-py3-none-any.whl") is true, while hosted_patch_uuid returns None. A real hosted sdist URL, or a path-prefixed hosted URL, returns false in the vendored check and Some(uuid) in hosted_patch_uuid. The refusal code is the same either way (pypi_pipenv_source_already_exists, #L229-L240),`` but the remedy it prints ("run socket-patch rollback" versus "user-authored") is wrong in both directions.

Symptoms

None filed.

Impact:

• Hosted Pipenv on a path-prefixed --patch-server-url deployment cannot re-run or rotate grants: every scan after the first refuses.
• Vendored Pipenv gives misleading remediation.
• The grammar will drift again with the next shape change. Poetry, pdm and uv use other recognizers; check them in the same PR.

Size: small.

Proposed change

• Make one recognizer the authority: hosted_artifact_url for the coordinates, combined with hosted_patch_uuid's origin allowlist (the configured origins plus patch.socket.dev) for "is it ours". Put it next to hosted_artifact_url, or in redirect::hosted_url.
• Hosted owned_url becomes "the shared recognizer accepts it, and its name, version and leaf match dep".
• Vendored is_socket_hosted_reference becomes "the shared recognizer accepts it with the configured origins".
• Delete both private grammars (the segment-count checks).

Size and scope

patch/redirect/pipenv.rs, vendor/pypi_pipenv.rs, and possibly vendor/lock_inventory/pypi.rs. About −40 / +25 production lines. Out of scope: the other pypi lock writers, beyond a note in the PR if their checks differ.

Acceptance criteria

• Neither private grammar remains.
• Regression test: hosted Pipenv rotates its pin on a path-prefixed --patch-server-url origin.
• Regression test: vendored Pipenv treats a foreign-host /patch/pypi/… URL as user-authored, and treats a hosted sdist or path-prefixed URL on an accepted origin as Socket-hosted.
• owned_url_follows_the_grant_origin, compatibility_tests, the vendored Pipenv tests and vex/discover/pypi_* stay green.

Dependencies

None. This is independent of E23 (the pypi backend skeleton) but shrinks it.