bug(packages): restore aborts on first failed install; migration report never written (re-files #46) #103

Summary

Re-files #46, which was closed on 2026-07-03 with no fix commit, no closing PR, and no comment (verified via the issue timeline — the closed event has a null commit_id). Both halves of the original finding are still present on main (e37b326):

1. Restore aborts on the first failed install. internal/packages/snapshot.go:171-179 — installPackages returns on the first npm install -g error, so one broken/renamed/yanked package kills the migration of every package after it. A user upgrading with 30 globals where feat(detector): implement Volta detection with mocked tests #3 fails loses 27 packages silently.

2. Migration report is never written. internal/packages/report.go (MigrationReport, PackageResult, Save()) has zero call sites anywhere in the repo — the entire file is dead code. platform.DataDir()/reports/ exists as a documented data dir but nothing ever writes to it.

How to fix

Part 1 — continue past failures, collect results

Rework installPackages to loop all packages, recording a PackageResult per package instead of returning early:

func installPackages(ctx context.Context, pkgs []Package) ([]PackageResult, error) {
	results := make([]PackageResult, 0, len(pkgs))
	var failed int
	for _, pkg := range pkgs {
		if ctx.Err() != nil {
			return results, ctx.Err() // Ctrl-C still aborts promptly
		}
		_, err := platform.RunShell(ctx, "npm", "install", "-g", pkgSpec(pkg))
		r := PackageResult{Name: pkg.Name, AttemptedVersion: pkg.Version, Status: "ok"}
		if err != nil {
			r.Status = "failed"
			r.Error = err.Error()
			failed++
		}
		results = append(results, r)
	}
	if failed > 0 {
		return results, fmt.Errorf("%d of %d packages failed to install", failed, len(pkgs))
	}
	return results, nil
}

Context cancellation must still abort the loop (the ctx.Err() check) — only per-package npm failures continue.

Part 2 — wire up the migration report

• Restore / RestoreFromSnapshot return the []PackageResult (or accept a callback) so the caller can build the report.
• In internal/cli/upgrade.go's restore step (~line 309-317), construct a MigrationReport{Timestamp, Manager, FromVersion (oldVersion), ToVersion (new default), Results} and Save() it to platform.DataDir()/reports/ — on both success and partial failure.
• Print the report path in the post-restore summary, especially when some packages failed, so the user knows what to re-install by hand.

Sentinel semantics (decide + document)

upgrade.go:100-103 removes the sentinel only when restoreSucceeded. With partial failures now possible, define: partial failure keeps the sentinel (restoreSucceeded = false), so the interrupted-upgrade replay path stays available — but the report still gets written either way. Document this in the sentinel arming comment.

Tests

• Table-driven test with the runShell seam failing on package N of M → asserts packages N+1..M were still attempted and the returned results carry the right statuses.
• Test that a report file lands in reports/ with correct Results on partial failure.
• Test that context cancellation aborts mid-loop (results truncated, error is ctx.Err()).

Acceptance criteria

• One failing global package no longer prevents the rest from migrating.
• Every nodeup upgrade migration writes a report file to platform.DataDir()/reports/; path surfaced to the user on partial failure.
• internal/packages/report.go has real call sites (no longer dead code).
• Sentinel behavior on partial failure is deliberate and documented.