v5 WS4/WS6: one hosted engine for disk + memory; unified Ledgers view

[Mikola Lysenko (mikolalysenko)]

Implements WS4 and a first cut of WS6 from docs/design/v5-plan.md, on top of #280 (ledger-free hosted, 686e5fb).

WS4: one hosted engine (disk + memory)

• The plan → rewrite stages of run_redirect_selected now live in socket_patch_core::hosted::engine, as pure functions over ProjectView: build_candidates, bun_lockb_symlinked, withhold_everywhere, read_candidate_files, wheel_targets, rewrite and guard.
• The disk flow (scan / get --mode hosted) keeps only the host side:
  • the apply lock
  • the probes: pipenv version, and stale installs for gem, python and vlt
  • the vendored takeover (vendored_takeover)
  • the symlink refusal
  • the commit of the rewritten files
• The in-memory engine moves to socket_patch_core::hosted::memory and runs the same stages over ProjectView::Memory. Its duplicated redirect orchestration is deleted.
• Other code that moved:
  • the pnpm trustLockfile / npm allow-remote planners → hosted::guidance
  • the vlt preflight → hosted::vlt
  • the JSON shaping of skipped patches, warnings and the redirect block → hosted::render; the engine emits typed RewriteWarnings, not serde_json::Value.
• No hosted ledger. Following v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280, neither engine reads or writes .socket/vendor/redirect-state.json. This branch's old hosted::ledger module (the redirect-ledger merge, in-memory load and serializer) is deleted, not moved.
• socket-patch-node now depends on socket-patch-core only. The CLI re-exports hosted_memory for hosted-bundle and the tests.
• hosted_memory_parity, redirect_golden and v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280's upstream_restore_golden pass on the merged tree.

WS6: one Ledgers view + ProjectContext

socket_patch_core::ledgers holds the one owner-precedence rule:

• manifest > vendor ledger > hosted records, by ledger key;
• a manifest key claims every vendor entry filed under it or naming it as base purl.

The hosted records are what v5 has instead of a ledger: the lockfiles' hosted pins, or a run's fetched records, plus a pre-v5 ledger that is read, never written, for migration.

The rule exposes these views:

• owned: one group per owner key, with the losing copies as alternates;
• listed: every copy worth showing;
• matching: store entries a remove/rollback identifier matches;
• hosted_vendored_overlap.

These views replace:

• list's manifest and vendor folding (hosted pins are listed one per pin, as in v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280);
• fold_vendor_records / vendor_record_is_unowned;
• the vendor half of scan's merge_ledger_records_for_updates (v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280's precedence is kept: manifest > hosted pins > vendor ledger);
• scan's overlap_from_states;
• vex_sources::build_candidates;
• the manifest and vendor matching loops in rollback and remove.

LoadedLedgers::load loads the manifest, the vendor ledger and any pre-v5 redirect ledger once, and each store keeps its own load outcome.

commands::context::ProjectContext loads the stores, the lockfile inventory and the wiring discovery lazily, at most once per run:

• locks() and discovery() read --cwd through one read-through DiskSnapshot (ProjectView::Snapshot), so each lock or config file is read at most once and both see the same bytes.
• Scan's discovery phase and hosted-pin reads, list and get go through it.

Behavior differences, listed in the CHANGELOG:

• scan's updates[] no longer folds a vendor entry the manifest claims by base purl.
• vex treats every vendor entry a manifest key claims as a fallback copy of that key's record.
• setup --check no longer folds a detached vendor entry whose base purl is in the manifest.
• get falls back to the committed artifacts on a corrupt vendor ledger, as scan does.

Follow-ups (not in this PR)

Listed in a PR comment: the rest of review item 3 (share the parsed models; thread the context into vendor / vex) and item 4 (a typed HostedOutcome with JSON, human and memory adapters), plus F17.

Validation (on 4cabaf1, the merge)

• cargo clippy --workspace --all-targets --all-features -- -D warnings is clean.
• cargo test --workspace --all-features --no-fail-fast -j4:
  • this branch: 10216 passed, 18 failed;
  • release/v5-prerelease @ 686e5fb, same sandbox: 10211 passed, 18 failed.
  • The 18 failures are the same tests on both, all root-user write-failure tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KHZ8uzdXfkG2zH8ZYDG8ju

---

Note

Medium Risk
Broad refactor of how patch stores are merged and how hosted rewrites run, with intentional behavior changes across scan, vex, get, and setup rather than cosmetic CLI moves.

Overview
This PR centralizes hosted-mode lockfile rewriting in socket_patch_core::hosted so the CLI disk path and the in-memory engine (Node addon / hosted-bundle) share the same plan → rewrite stages over ProjectView. The CLI keeps host-only concerns (apply lock, vendored takeover, symlink guards, commits); pnpm/npm install-policy text, vlt preflight, and JSON rendering move into core. socket-patch-node no longer depends on the CLI—it uses core only, with hosted_memory re-exported for compatibility.

It also introduces socket_patch_core::ledgers with a single precedence rule (manifest → vendor ledger → hosted records, including manifest claims by ledger key or base purl) and ProjectContext to load stores, lock inventory, and wiring discovery once per run (shared DiskSnapshot for lock/config reads).

Commands list, scan, vex, get, setup --check, rollback, and remove now read patch stores through that view instead of ad hoc merging. User-visible fixes include: scan updates[] no longer double-counts vendor entries the manifest owns by base purl; vex treats claimed vendor rows as fallback copies of the manifest record; setup --check stops folding detached vendor entries when the manifest already records the base purl; get matches scan when the vendor ledger is corrupt (fallback to committed artifacts). Manifest record building from API patches moves from get into core for reuse.

^{Reviewed by Cursor Bugbot for commit 4cabaf1. Configure here.}

[Mikola Lysenko (mikolalysenko)]

vlt patch compatibility / install-proof (ubuntu-latest, *) fails on vlt_pinned_matrix_agent_get_and_remove with left: Absent, right: Patched at tests/e2e_vlt.rs:156.

This is not caused by this PR. The same test fails the same way on release/v5-prerelease @ 8ae7dc3 (for example run 36352437746, job install-proof (ubuntu-latest, 0.0.0-18)).

Root cause: v5 get defaults to hosted mode (commands/get.rs: args.mode.unwrap_or(... ScanMode::Hosted)). This agent-mode test still calls get <uuid> without a mode, so it now runs the hosted rewrite plus the vlt warm-tree heal, which removes the stale installed copy. That is why the package reads Absent instead of Patched.

No fix exists yet. Proposed patch, a test-only change I'm not adding here to keep this PR's scope:

--- a/crates/socket-patch-cli/tests/e2e_vlt.rs
+++ b/crates/socket-patch-cli/tests/e2e_vlt.rs
@@ async fn vlt_pinned_matrix_agent_get_and_remove() {
-    let out = socket_api(&fx.proj, &fx.svc, &["get", UUID], &[]);
+    let out = socket_api(&fx.proj, &fx.svc, &["get", UUID, "--mode", "agent"], &[]);

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

CI / coverage (and test (*) / test-release) fail on e2e_redirect_cargo_build::cargo_get_uuid_hosted_fresh_checkout_fetch and cargo_hosted_fresh_checkout_fetch_pulls_patched_crate_and_vex_verifies. The failing assertion is (3) offline: left Some(0), right Some(1): an offline vex still attests the patch (redirected).

This isn't from this PR. release/v5-prerelease @ 8ae7dc3 fails the same two tests with the same assertion (CI run 36352437716, job coverage), and they fail locally on the base branch as well.

Likely cause, not yet verified: step (2) runs the "read-only" embedded scan --vex with no --mode. In v5 that defaults to hosted, so it writes the redirect ledger and its records. Step (3) ("offline with no ledger") then finds the record locally. A fix is probably to run that step as scan --mode agent --vex … or --dry-run, or to delete .socket/vendor/redirect-state.json before (3). No fix exists on the base branch yet, so I'm not changing it in this PR.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

CI / test (windows-latest) fails two targets: covgap_commands_scan_mod and e2e_redirect_cargo_build. The cargo one is covered in my earlier comment. release/v5-prerelease @ 8ae7dc3 fails exactly the same two targets on Windows (CI run 36352437716, job test (windows-latest)), so neither failure comes from this PR. The Linux and macOS test jobs fail only the cargo target. No fix exists on the base branch yet.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

v5 review at 5dcf9882 — moving the hosted engine into core and removing the Node addon's CLI dependency is the right boundary. Keep it. I did not confirm a new functional regression, but there is avoidable work and unfinished integration.

1. Index ledger ownership once. Ledgers::owned() sorts and scans all vendor keys inside each manifest-record iteration (ledgers.rs:173–192). For M manifest and V vendor entries this adds O(M·V log V) work, even for direct key matches. Build a deterministic owner-to-entry index once and derive the different views from it. This is a source-level complexity finding, not a measured wall-clock regression.
2. Finish the ledger-free integration before calling this complete. Disk and memory still serialize the hosted ledger, and the new abstraction models three persistent stores. With #280, retain only legacy read/migration support for hosted records; remove the hosted writer/merge path from both engines. The #280/#282 heads conflict across these paths, so parity must be checked again on the combined tree.
3. Cache one project snapshot, not two interpretations of the files. ProjectContext::locks() and discovery() independently read/parse the project, and the context has no shared crawl snapshot. Back both with the shared format models and a consistent project root; rebuild/invalidate the snapshot after writes before embedded VEX. Thread it through vendor/VEX as well as scan/get/list.
4. Keep rendering out of orchestration. Return a typed plan/result from the shared engine, then let disk, memory, JSON and human adapters consume it. This makes the reduction durable rather than relocating the old CLI state machine into core.

Validation: source/diff review; existing parity/golden coverage inspected, not rerun here.

[Mikola Lysenko (mikolalysenko)]

Thanks, replies by item:

1. Ledger ownership index: fixed in 04b3d96, not pushed yet; it's waiting on the full test run. Ledgers::claims() groups the vendor entries under the manifest key that claims them, in one pass over the sorted ledger. owned()/listed() read from that group, which removes the per-manifest-key re-sort and scan. Each view is now O(V log V + M). Ordering and alternates are unchanged: the ledgers unit tests and the 844 CLI lib tests pass unchanged, and clippy is clean.

2. Ledger-free integration. Agreed that this isn't complete until the hosted writer and merge path are gone. v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280 is still a draft with WS1 items open, and our heads overlap on hosted/engine.rs, scan/hosted.rs and the rollback/remove hosted legs. So my proposal is:

   • (a) keep v5 WS4/WS6: one hosted engine for disk + memory; unified Ledgers view #282 as the refactor with behavior parity, rebasing it onto release/v5-prerelease once v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280 lands;
   • (b) in that merge, delete hosted::ledger (delta + save) and the engine's LedgerEdits output from both the disk and memory paths;
   • (c) make Store::Hosted a read-only legacy source, used only by migration, list and vex;
   • (d) re-run hosted_memory_parity and redirect_golden on the combined tree.

   If you'd rather land v5 WS4/WS6: one hosted engine for disk + memory; unified Ledgers view #282 first and have v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280 rebase onto it, say so. I kept the ledger delta isolated in hosted::ledger (the engine never calls it) so that either order is a deletion, not a rewrite.

3. One project snapshot. Planned next on this branch:

   • ProjectContext gets one ProjectSnapshot: the root plus the parsed lock/config files through the shared format models. locks() and discovery() both derive from it instead of reading files twice.
   • A write through the engine commit returns the touched paths, and the context invalidates them. Embedded scan --vex then re-derives from the refreshed snapshot rather than re-crawling.
   • Thread the context through vendor and vex too.

4. Rendering out of orchestration. Also planned here. The engine will return a typed HostedOutcome: confirmed, skipped, the warnings as typed variants, files, and the stale-install results. It will no longer build serde_json::Value warnings. The JSON envelope and the human output become adapters over it, as will the memory engine's JS result. redirect_json_block / format_* move into a render module that only consumes the outcome.

I'll push 1 once the suite is clean, then do 3 and 4 as separate commits on this PR. Item 2 waits on the #280 ordering decision.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Status on the review as of f39ed1c:

1. Ownership index: done (04b3d96). Ledgers::claims() builds the owner → entries index in one pass. owned() and listed() both derive from it.
2. Ledger-free integration: still waiting on the v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280 ordering decision. The plan is in my earlier reply.
3. One project snapshot: partly done (f39ed1c).
   • Done:
     • ProjectView::Snapshot wraps a read-through DiskSnapshot, so each lock or config file is read at most once per run and every reader sees the same bytes.
     • ProjectContext::locks() and discovery() both read --cwd through that one snapshot.
     • Lockfile discovery's guarded reads now go through a ProjectView (discover_patched_refs_in).
   • Remaining:
     • Share the parsed format models, not only the bytes.
     • Invalidate the snapshot after the engine's writes so embedded --vex reuses it.
     • Thread the context into vendor and vex.
4. Rendering out of orchestration: partly done (f4e8d54).
   • Done:
     • The engine emits typed RewriteWarnings instead of serde_json::Value. This covers rush/pnpm/npm guidance, the vlt preflight, fetch failures and the memory engine's pre-warnings.
     • The JSON shaping lives in hosted::render: skipped_json, rewrite_warnings_json, redirect_json_block.
   • Remaining: a single typed HostedOutcome returned by the engine, with the disk JSON, human output and the memory JS result as adapters over it.

CI on f39ed1c: every red check is one of the base-branch failures I commented on above (vlt agent_get_and_remove, cargo fresh_checkout_fetch, Windows covgap_commands_scan_mod). The earlier Bun production and Poetry failures went green on this head. Locally the full suite gives 10320 passed; its only failures are the ones the base branch also has in this sandbox. Clippy -D warnings is clean.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Restack request from the v5 coordinator (please act now).

#280 (ledger-free hosted) is now merged into release/v5-prerelease as 686e5fb4. Please merge origin/release/v5-prerelease into this branch, resolve the conflicts, get CI to "base-inherited reds only", and mark the PR ready for review (not draft) when done.

• Resolve toward v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280's model. There is no hosted ledger any more: rollback/remove/vendor restore upstream via patch::redirect::upstream. Don't reintroduce redirect-state.json writes or fragment-replay paths. Keep this PR's own intent on top.
• Landing order: WS5: one VendoredBackend for vendored apply/revert/repair; cut repair's ledger rebuild #283 → WS3: one lockfile model per ecosystem #281 → v5: remove setup (WS7) + patch UI streamlining (WS8) #279 → v5 WS4/WS6: one hosted engine for disk + memory; unified Ledgers view #282, then the design PRs (docs: v5 waste review + repacking-to-depscan design #286 docs, Add the NuGet vendoring design (docs only) #285/Vendored Maven and Gradle: design and prototype #287). After each one lands, the next ones merge the base again. Expect a small second round of conflicts.
• Base-inherited reds that aren't yours to fix: vlt install-proof (vlt_pinned_matrix_agent_get_and_remove; fixed on v5: remove setup (WS7) + patch UI streamlining (WS8) #279) and Windows covgap_commands_scan_mod.
• The repo only allows squash merges. Post a comment with your final head SHA and CI summary when you're ready.

[Mikola Lysenko (mikolalysenko)]

Owner decision (via coordinator): on ordering, #280 has landed first (686e5fb4). Merge the base, drop the hosted-ledger writer/merge from both engines (legacy read only), and keep parity green. Items 3 (shared parsed snapshot) and 4 (typed HostedOutcome) become follow-up PRs. Priority is landing after #279.

[Mikola Lysenko (mikolalysenko)]

I merged release/v5-prerelease @ 686e5fb (#280) into this branch at 4cabaf1.

Resolved toward #280's ledger-free model:

• socket_patch_core::hosted::ledger is deleted, not moved (docs: v5 waste review + repacking-to-depscan design #286 F26).
• Neither the disk engine nor the memory engine reads or writes redirect-state.json, and nothing replays fragments.
• This PR still does the one-engine unification and applies the shared owner rule to the manifest and vendor ledger.
• The hosted records are now the lockfiles' hosted pins. A pre-v5 ledger is read only, for list details and vex.
• updates[] keeps v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280's precedence: manifest, then hosted pins, then vendor ledger.

Local validation on the merge:

• clippy with -D warnings is clean.
• The full suite has the same 18 failures as the base in this sandbox, all root-user write-failure tests: 10216 passed here, 10211 on the base.
• The parity and golden suites pass: hosted_memory_parity, redirect_golden, upstream_restore_golden.

I'll mark the PR ready once CI on 4cabaf1 shows only base-inherited reds, and I'll re-merge the base after #283, #281 and #279 land.

Follow-up list (non-blocking, not in this PR):

1. Review item 3, rest: share the parsed format models across ProjectContext::locks() and discovery() (today they share bytes, not parses). Invalidate the snapshot after the engine's writes so embedded --vex reuses it. Thread the context into vendor and vex.
2. Review item 4, rest: a single typed HostedOutcome from the engine, with disk JSON, human output and the memory JS result as adapters over it.
3. Move the remaining single-store load sites (vendor, repair, apply, bun preflight) onto ProjectContext.
4. F17: the napi addon (socket-patch-node) and hosted-bundle have no consumers. Listed only, no action here.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

This PR is ready for review at head 4cabaf1d3193bcf7ab6d2ae2c5255bb08b58c7e0, which merges release/v5-prerelease @ 686e5fb (#280).

CI on 4cabaf1:

• CI: 38 jobs passed and 5 were skipped. The only red is test (windows-latest), and the only failing target in it is covgap_commands_scan_mod, which also fails on the base branch.
  • The first attempt was cancelled at the 50-minute job budget with no failures. The re-run finished in time.
• vlt patch compatibility: every install-proof leg fails only vlt_pinned_matrix_agent_get_and_remove. That failure also comes from the base branch, and v5: remove setup (WS7) + patch UI streamlining (WS8) #279 fixes it.
• Bun patch compatibility: green on re-run.
  • The first attempt failed only three macOS 1.3.10 workspace-hosted legs, where bun install --frozen-lockfile timed out after 180s.
  • Their vendored twins, and every other leg, passed.
• Other compatibility workflows: pnpm, npm, Poetry, Pipenv, PDM and Go are green.

Local run on the merge:

• clippy with -D warnings is clean.
• The full suite has the same 18 failures as the base in this sandbox, all root-user write-failure tests.

I'll merge the base again after each of #283, #281 and #279 lands.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

#283 landed on release/v5-prerelease as 06437d2; please merge origin/release/v5-prerelease again, resolve conflicts, get green, and keep it ready.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

This PR is at head 907de9c, which merges release/v5-prerelease @ 06437d2 (#283). The merge was clean, with no conflicts, and the PR is still ready for review.

CI on 907de9c:

• CI: 38 jobs passed and 5 were skipped. The only red is Windows covgap_commands_scan_mod, which also fails on the base branch.
• vlt: the install-proof legs fail only agent_get_and_remove, which also fails on the base branch and is fixed on v5: remove setup (WS7) + patch UI streamlining (WS8) #279.
• Bun: green on re-run.
  • The first attempt hit a DNS failure on a macOS 0.8.1 runner (nodename nor servname provided).
  • It also hit a timed-out rollback in one macOS 1.3.0 leg.
• Other workflows: pnpm, npm, Poetry, Pipenv, PDM and Go are green.

Local checks on the merge:

• clippy with -D warnings is clean.
• The full suite has the same 18 failures as 06437d2 in this sandbox, all root-user write-failure tests: 10168 passed here, 10163 on the base.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

#281 landed on release/v5-prerelease as 73c0c4f; please merge origin/release/v5-prerelease again, resolve conflicts, get green, and keep it ready.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

CI on 94dc5d2 (the merge of #281 base 73c0c4f). Two reds so far:

1. Pipenv compatibility / matrix (ubuntu-latest, 2018.11.26 … 2026.8.0): the hosted legs failed at random across versions (dryRunParity, appliedExactlyOne, and so on). macOS passed on the same versions. The single re-run passed with no change, so it is transient.

2. vlt install-proof ubuntu 0.0.0-11 and ubuntu 0.0.0-1: e2e_redirect_vlt_build fails vlt_pinned_matrix_hosted_{rollback_byte_exact,crlf_lock,idempotence}. Hosted rollback doesn't restore the node tuple's 4th element, the original https://registry.npmjs.org/…tgz URL. It either drops it or rewrites it to the options.registry host. This is not this PR's failure. With vlt 0.0.0-11 locally, the same 3 tests fail identically (35 pass / 3 fail) on:

   • the base 73c0c4f
   • 28cebf7, before WS3: one lockfile model per ecosystem #281
   • this head

   It comes from the ledger-free upstream restore for vlt pins (v5 WS1+WS2: ledger-free hosted mode, upstream-restore rollback, vendor ejects hosted projects #280 model, patch/redirect/upstream/vlt.rs). That restore rebuilds the tarball URL from the lock's registry, so a pre-alpha vlt lock that carried an explicit npmjs URL under a different options.registry can't round-trip. No fix exists on the base or in an open PR, so nothing is ported. Suggested fix, as a follow-up on the base: when the pin was written, keep the replaced tuple URL in the pin itself (as the uv/pypi pins keep their source), and have upstream::vlt restore that instead of re-deriving it. The alternative is to scope these three legs' byte-exact assertion to vlt ≥ 1.0 locks.

The rest of CI is still running. I'll post the head SHA and CI summary when it finishes.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

#279 landed on release/v5-prerelease as f6bdad5; please merge origin/release/v5-prerelease again, resolve conflicts, get green, and keep it ready.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

v5 coordinator: #282 is next to land and is up to date with the base (f6bdad5, #279). It has no unresolved threads and is ready for review. Only one thing blocks landing: CI run 36474481876 on head 24a8241 is still in progress.

vlt patch compatibility failed on this head. Three install-proof legs are red: ubuntu 0.0.0-1, ubuntu 0.0.0-1 on Node 22.0.0, and ubuntu 0.0.0-11. The windows 0.0.0-11 leg is red too. When CI finishes, please post the head SHA and CI summary. For each red, show that it is base-inherited: it fails on the base's latest CI run under the same job name, or it is Windows covgap_commands_scan_mod. The coordinator will land this PR on its next run once that holds.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Head 24a82415f506327072d04c793c10c788ae40268b: the merge of release/v5-prerelease at f6bdad5 (#279). The PR is up to date with the base and ready.

Local validation before the push:

• clippy -D warnings is clean.
• The full workspace suite has the same failure set as the base f6bdad5: 12 failures on each (root-sandbox write-failure tests), with 9435 passing on the merge vs 9430 on the base.

CI on this head

CI run 36474481876: 228 of 233 jobs pass. The Windows test job finished inside its timeout. The 5 reds:

Job	Base evidence	Why it is not this PR's
yarn-classic 1.0.2, 1.6.0, 1.7.0, 1.9.4	The same four job names are the only reds on page 1 of #279's final CI run, 36466224860 (head 88c444c, squash-landed as f6bdad5).	They are base-inherited: mode_migration_npm in the e2e tier enabled by #288.
e2e-docker (deno)	Passes on 36466224860.	Infrastructure: curl https://deno.land/install.sh exited 35 (TLS connect) while building the Docker image, before any test ran. I re-ran the failed jobs once just now.

vlt patch compatibility run 36474482017 has 4 reds:

• install-proof (ubuntu-latest, 0.0.0-11)
• install-proof (ubuntu-latest, 0.0.0-1)
• install-proof (windows-latest, 0.0.0-11)
• install-proof (ubuntu-latest, 0.0.0-1, 22.0.0)

These four are exactly the failed jobs of #279's final vlt run, 36466226104 (head 88c444c). That workflow runs only on pull_request, so the base branch has no vlt run of its own and #279's final head is the closest base run. Each leg fails vlt_pinned_matrix_hosted_{crlf_lock,idempotence,rollback_byte_exact}, and those three tests fail identically locally on 73c0c4f and 28cebf7. Details are in my earlier comment.

Every other workflow on this head is green: pnpm, Poetry, Go, Pipenv, npm, PDM and Bun.

---

Generated by Claude Code