Fix Cargo source replacement being ignored (#338, #455)

[Mikola Lysenko (mikolalysenko)]

LLM Description written by Claude Code:claude-opus-5-5

Fixes #338
Fixes #455

Summary

socket-patch never read Cargo's source replacement
([source.crates-io] replace-with = "…" → [source.<name>] directory = "…"),
so it guessed where a cargo vendor project builds from, and guessed wrong.

• Agent mode (Agent-mode cargo apply patches the wrong copy when cargo vendor uses a custom directory (or apply runs from a workspace member), yet reports success and VEX attests not_affected #338, Hosted cargo scan in a cargo vendor project reports success but breaks every fresh cargo build --frozen --offline, and VEX then omits the patch #455 symptom 2). The crawler hard-coded
  <cwd>/vendor. With a custom vendor directory, with a workspace-root
  vendor directory seen from a member, or with an unwired or non-cargo
  vendor/, apply patched a copy cargo doesn't build from and reported
  success, and vex judged the wrong copy. The crawler now follows the
  replace-with chain of every config cargo reads (each ancestor's
  .cargo/config*, then $CARGO_HOME's) and crawls the directory source it
  ends on, resolved the way cargo resolves config paths (against the parent
  of .cargo/). Without a directory replacement it crawls the registry
  cache. A configured directory that is missing yields nothing, never a copy
  cargo ignores.
• Rollback/prune (Cargo rollback after an agent→vendored takeover leaves the shared registry cache patched, reports success, and deletes the revert blobs #336 leg). find_cargo_copies decided that a vendor
  directory hides the registry cache by checking <cwd>/vendor. It now asks
  the crawler for the configured directory source.
• Hosted mode (Hosted cargo scan in a cargo vendor project reports success but breaks every fresh cargo build --frozen --offline, and VEX then omits the patch #455 symptom 1). When the project's .cargo/config*
  replaces crates.io with a directory source, every cargo dependency is
  skipped with nothing written and a new additive
  redirect_cargo_directory_source warning that points at agent mode. A
  registry-mirror replacement is still redirected.

The new pure module formats/cargo/source_replacement.rs holds the resolver
(crates_io_source) and the config file walk (config_layer_paths).
CLI_CONTRACT.md and docs/ecosystems.md are updated.

Per-issue checklist

Issue	Regression tests (red before the fix → green)
#338 shape 1 (custom dir)	crawler_cargo_e2e::custom_vendor_dir_named_by_source_replacement_is_crawled; real cargo: e2e_safety_cargo_build::apply_patches_a_custom_named_vendor_dir
#338 shape 2 (member)	crawler_cargo_e2e::workspace_root_source_replacement_applies_to_a_member; real cargo: e2e_safety_cargo_build::apply_from_a_member_patches_the_workspace_vendor_dir
#338 shape 3 + non-cargo vendor/	crawler_cargo_e2e::unwired_vendor_dir_falls_back_to_the_registry_cache
#338 fresh checkout / missing dir	crawler_cargo_e2e::missing_replacement_directory_crawls_nothing
#338 $CARGO_HOME config	crawler_cargo_e2e::cargo_home_source_replacement_is_honoured
#455 hosted redirect	patch::redirect::tests::cargo_directory_source_replacement_is_refused (both config spellings, CRLF, a multi-hop chain; mirror control still redirects)
#455 vex reads the wrong copy	covered by the crawler change: vex discovers through the same crawler (vex_consumed.rs); the crawler tests above
resolver	formats::cargo::source_replacement::tests (8 unit tests)

Red before the fix: all 5 new crawler_cargo_e2e tests failed on origin/main
code. The hosted test failed with the Socket registry pin written into
Cargo.toml, Cargo.lock and .cargo/config.toml.

Fixtures that staged a bare vendor/ and relied on the old guess now write the
.cargo/config.toml that cargo vendor asks for (6 crawler unit/oracle tests,
crawlers_empty_paths_e2e, e2e_cargo ×2, e2e_vendor_cargo_build's #336
"behind a cargo vendor dir" rollback). None of their assertions were loosened.

Local evidence (on 30aa879; 9083269 adds 3 fixture wirings, CI green)

• cargo fmt --all -- --check: clean
• cargo clippy --workspace --all-features -- -D warnings: clean
• cargo test -p socket-patch-core --all-features: all pass except 4
  read-only-permission tests that can't fail when run as root in this sandbox
  (copy_tree::relax_loop_must_not_traverse_symlinked_root,
  vlt_heal::an_unremovable_hidden_lock_keeps_every_store_entry,
  pypi_poetry::wire_write_failure_…, pypi_requirements::wire_failure_…).
  These are unrelated to this diff.
• socket-patch-cli: --lib, e2e_cargo, in_process_cargo_apply,
  mode_migration_cargo, e2e_redirect_cargo_build, e2e_redirect_cargo_shapes,
  e2e_vendor_cargo_build (incl. both Cargo rollback after an agent→vendored takeover leaves the shared registry cache patched, reports success, and deletes the revert blobs #336 rollback tests with real cargo),
  vex_terminal_output, in_process_rollback_hosted, in_process_vendor,
  covgap_commands_scan_mod: pass
• cargo test -p socket-patch-cli --all-features --test e2e_safety_cargo_build -- --ignored:
  10/10 pass with real cargo 1.93.1, including the two new Agent-mode cargo apply patches the wrong copy when cargo vendor uses a custom directory (or apply runs from a workspace member), yet reports success and VEX attests not_affected #338 shapes
• The full cargo test --workspace couldn't finish locally: the sandbox ran
  out of disk building every CLI test binary. CI runs it.
• Wrappers (npm/, pypi/, gem/) only dispatch to the binary, so no
  parallel change is needed.

/code-review high: findings not fixed here

• Hosted mode reads only the project's own .cargo/config* (the
  rewriter is pure over the project's files). A directory replacement set
  only in an ancestor or ~/.cargo/config.toml is not seen by hosted mode.
  Hosted cargo scan in a cargo vendor project reports success but breaks every fresh cargo build --frozen --offline, and VEX then omits the patch #455's repro and cargo vendor's printed snippet both put it in the
  project config. Plumbing outer configs into the hosted engine (like
  yarn's OuterYarnMirror) is a follow-up.
• Vendored mode's is_vendored (vendor/cargo.rs) still probes
  <root>/vendor for its already_vendored_in_tree refusal. That is a
  fail-closed refusal with its own contract tests, and neither issue
  reports it. Follow-up.
• Upgrade from a release that patched an unwired vendor/. Cargo never
  built from that copy, so the patch had no effect. rollback now looks
  where cargo builds. The stray edit in the unused vendor/ copy stays.
• local-registry / git replacements are Other and keep the
  registry-cache crawl. Cargo unpacks local-registry crates into
  registry/src, which is what the crawler reads.
• Lexical .. in --cwd and per-key (not per-table) merge of
  conflicting [source.x] definitions: cargo itself rejects the latter
  ("more than one source location"). absolute() is used rather than
  canonicalize() so reported paths keep matching the cwd the user gave
  (macOS /var → /private/var).
• A third config-chain walker / no memo. vendor::cargo_config::read_config_chain
  is [patch]-specific (it drops files with no patch entries). Unifying the
  walkers is a refactor for its own PR. The extra reads are at most a handful
  of small files per crawl.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TUHnkhA7W5gHh21jSMkY7P

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 9083269. Configure here.}