GeezerVison

Lane ADE of the yarn mode-conversion campaign: three adversarially-verified SILENCE gaps, each pattern-fixed with additive warnings only (precedent: #133/#159/#164/#189/#201). No exit-code or status changes anywhere. CLI_CONTRACT.md updated for all three.

Root causes (file:line, pre-fix)

A — hosted→agent is a silent no-op (crates/socket-patch-cli/src/commands/scan/mod.rs, agent branch ~L1545 JSON + human path)
scan --mode agent over a live hosted redirect reported status: success / applied with zero warnings while yarn.lock stayed pinned to patch.socket.dev and .socket/vendor/redirect-state.json stayed live. The takeover machinery structurally cannot see it: classify_overlap_takeover (mod.rs:602) keys on purls present in both ledgers (hosted ∩ vendored), so hosted-only wiring never triggers it — and the agent path loaded the redirect ledger (mod.rs:1472) only for updates[] merging. No npm/yarn hosted revert exists, so the "conversion" can never complete without another mode run.

D — vendored→agent refusal invisible at the envelope (mod.rs:1564-1571 via partition_skipped_selected(..., "vendored"), vendor_flow.rs:503-549)
Agent-mode --json recorded vendor-owned skips only deep in apply.patches[] (skipped/vendored); no top-level signal, no stderr line — a JSON consumer had to dig per-patch to learn the mode did not change. (Human path already printed [skip] … (vendored …).)

E — berry rewriter misdiagnoses our own vendored entry (crates/socket-patch-core/src/patch/redirect/mod.rs:2039-2053)
The unsupported-protocol branch refused socket-patch's OWN file:./.socket/vendor/… wiring with a detail hardcoding (workspace:/patch:/portal:/link:) — file: not even listed — and named no remedy.

Matrix evidence

Repro'd by the matrix sweep (cells convy1-hosted2agent, convy4-hosted-agent, edgey1-roundtrip; convy1/convy4-vendored2agent; convy4-vendored-hosted) and re-verified here against production on throwaway copies of the verify dirs:

• A old binary: exit 0, status success, warnings: null, grep -c patch.socket.dev yarn.lock = 1 after "conversion". New: same exit/status, warnings[0].code = hosted_wiring_retained (JSON + stderr), state untouched.
• D old: exit 0, warnings: null, skip only in apply.patches[]. New: same envelope + warnings[0].code = vendored_ownership_retained naming the purl and remove <purl> / scoped vendor --revert migration path; vendored wiring untouched.
• E old: redirect_yarn_berry_unsupported_protocol claiming (workspace:/patch:/portal:/link:). New: redirect_yarn_berry_vendored_entry naming the retirement path (remove <purl> / scoped vendor --revert, then re-run scan --mode hosted); other protocols now get the generic code with the entry's ACTUAL protocol. Lock byte-identical, exit 0, both binaries.

Fix summary

• New run-level warnings[] (top-level {code, detail}, additive, mirrored to stderr unless --silent) on the agent apply flow:
  • hosted_wiring_retained — new hosted_wiring_retained_purls probes the already-loaded redirect ledger, reusing hosted_wiring_live per scanned purl (deliberately NOT via classify_overlap_takeover, which requires a vendored-ledger intersection). Fires only when ledger records are still live at scan time AND the live lock proves hosted wiring — so lane B's hosted→vendored pre-revert (which retires the records) retires this warning with them, pinned by scan_agent_hosted_warning_silent_once_ledger_records_are_gone. Detail: stay --mode hosted or migrate via scan --mode vendored; never a hosted→agent unwind (none exists), never hand-deleting the ledger.
  • vendored_ownership_retained — fires when the vendored partition is non-empty, naming the purls + remove <purl> (or vendor --revert, blast radius called out), per-patch records untouched.
• Berry rewriter: file: + .socket/vendor/ signature ⇒ distinct redirect_yarn_berry_vendored_entry code with the real remediation; otherwise the generic detail names the ACTUAL protocol(s). Refusal stays fail-closed/byte-identical.

Tests (TDD, RED commit 0842aca first)

• scan_invariants.rs: 4 new e2e cases (wiremock, seeded ledgers — no real yarn needed): D warning trigger; A warning trigger; A silence when records are gone (lane-B coordination guard); A silence on a registry-clean lock.
• scan/mod.rs unit tests beside the takeover suite: hosted_only_wiring_fires_agent_probe_not_the_overlap_classifier (pins the structural blindness), per-case silence guards, detail-wording assertions, distinct-code assertions.
• core redirect/mod.rs: vendored-entry code+remediation; actual-protocol naming for portal: and non-vendored file:; existing yarn_layering incident guard 4d still green.
• Gates: cargo test --workspace green, cargo clippy --workspace --all-features -- -D warnings clean, touched files rustfmt-clean.

Sibling-PR conflict notes (parallel lanes)

• Lane B fix/npm-vendored-over-hosted-prerevert (cli vendor.rs, scan/hosted.rs, core redirect takeover): no shared regions expected; this PR's A warning is coordination-safe with B by design (keys on ledger records still live at scan time — tested), and the E remedy wording holds in both worlds. If B touches scan/mod.rs near the takeover block, trivial context conflicts possible.
• Lane C fix/vendor-revert-driftskip-keep-artifacts (core vendor lock backends, vendor/mod.rs, cli vendor.rs revert path): disjoint files.
• PR fix(scan): surface yarn PnP refusal instead of silent success-0 across all modes #201 (scan/mod.rs empty-discovery path): same file, different region — possible context-level merge conflict in scan/mod.rs; semantics independent.
• PR test(yarn): version-matrix e2e capstones — yarn2/3 refusal pins, berry workspaces + pnpm-linker install proofs #200 (new e2e_yarn* files): disjoint.

🤖 Generated with Claude Code