sec(install): make cosign verification mandatory on Windows

[LukasWodka]

Closes tracebloc/backend#2078.

The bug

scripts/install.ps1 gated cosign verification on whether cosign happened to be installed:

if (Get-Command cosign -ErrorAction SilentlyContinue) { ...verify... }
else { Write-Host "  (cosign not installed; SHA256 verified, signature skipped)" }

On a fresh Windows box cosign is never present, so the default path was the unverified one. The SHA256 it kept is same-channel: SHA256SUMS ships from the same GitHub release as the binary, so whoever could swap the binary could swap the sums with it. It proves the download completed, not who built it.

install.sh has bootstrapped a pinned cosign and failed closed since backend#889. Windows was the one platform left, which made this claim in README.md:75 false — and false with no platform qualifier:

Signature verification is mandatory … fails closed

No README edit in this PR. The claim was correct about intent and wrong about Windows; this makes it true.

The fix

Windows now does what the other two platforms do — bootstrap a pinned, checksum-verified cosign, refuse to install when it can't. TRACEBLOC_ALLOW_UNVERIFIED=1 is the one escape, and it covers "cannot verify" only: no cosign, no .sig/.cert. It deliberately does not cover a verification that ran and failed, which is tampering evidence no env var should override. (AST-asserted, test 5.)

Two things this needs that install.sh does not:

The bootstrap fetches cosign-windows-amd64.exe on both architectures. Sigstore has never published a Windows arm64 cosign — at any release. Asking for a per-arch name 404s and locks Windows-on-ARM out permanently; that is the same bug as tracebloc/client#734, fixed there the same way. Verification is over bytes, so the verifier's instruction set cannot change the verdict, and the artifact handed to it is still the native arm64 binary.

Test-CosignRuns separates "cosign won't start here" from "the signature is bad." They arrive through the same channel and warrant opposite reactions — only one means the artifact may be tampered with. Absent x64 emulation on Windows-on-ARM is the case that makes it real, and "install cosign" would be useless advice there since one is installed.

Also sets a TLS 1.2 floor. PS 5.1 defaults to SSL3/TLS1.0 on older Windows, and every fetch here carries either the binary we are about to run or the verifier that authenticates it.

Tests

scripts/tests/install-ps1-verify.sh — install-verify.sh's sibling, same CI job.

	result
against develop's installer	0 passed, 5 failed (exit 1)
after this change	5 passed, 0 failed (exit 0)

install.ps1 can't be driven end-to-end on the Linux runner — it ends in Windows-registry PATH writes that throw there, so a full run would fail for reasons unrelated to what's under test. The behavioural tier therefore extracts the helpers from the real file by AST and executes them. A hand-copied helper would pass while production broke, which is the failure mode CLAUDE.md rule 9 names.

If pwsh is missing the harness fails rather than skips: "cannot tell" is not evidence that verification is mandatory (rule 3). ALLOW_NO_PWSH=1 downgrades it on a dev box; CI never sets it.

Six guards, each mutation-proven, each mutation asserted to have applied:

mutation	reddens
[bool]$env:… instead of -eq '1'	4 tests
checksum mismatch stops refusing	1
$LASTEXITCODE not armed before the probe	1
an opt-out on a FAILED verification	1
the no-cosign degrade restored	1
a \$ escape in a message	1

Three of those are bugs this change made, caught before review

• $AllowUnverified was [bool]$env:TRACEBLOC_ALLOW_UNVERIFIED. Every non-empty string casts to $true in PowerShell, so TRACEBLOC_ALLOW_UNVERIFIED=0 would have switched the bypass on. Now -eq '1', matching install.sh, with a truth table over unset / '' / 0 / false / no / true / 1.
• The bootstrap restated Get-Arch's logic minus its PROCESSOR_ARCHITEW6432 handling, so a 32-bit PowerShell host on ARM would have refused. Removed rather than fixed: the asset is arch-independent, so there was nothing to branch on.
• "Pin a signed \$env:RELEASE_VERSION" rendered as Pin a signed , — PowerShell escapes with a backtick, so \$ emits a backslash and expands the variable, dropping the one thing the user needed. Now a check of its own over every double-quoted message.

One test was vacuous on its first writing

A "stale exit code must not leak through" assertion built on an absent path — which throws and returns before ever reading $LASTEXITCODE. The mutation caught it: M3 applied and nothing reddened. Replaced with the input that is actually reachable — a PowerShell shim (scoop and chocolatey install tools as cosign.ps1, and Get-Command finds .ps1 on PATH whatever PATHEXT says), which & dispatches in-process and which sets no $LASTEXITCODE at all. Without the 255 preset the helper then reads the 0 left by the last successful command and reports a shim that did nothing as a working cosign.

Local

install-ps1-verify: 5 passed, 0 failed
install-ps1-functions: 22 passed, 0 failed
install-verify: 33 passed, 0 failed          # unchanged, no regression

shellcheck --severity=error clean on the new harness. actionlint on build.yml reports the same two pre-existing SC2001 style findings as HEAD (lines 132/145, unrelated).

Not in scope

TRACEBLOC_ALLOW_UNVERIFIED still has no Windows-side test that drives the installer end-to-end — only its helpers and its control-flow shape. Doing that properly needs a windows-latest job, which is a separate call about CI minutes.

🤖 Generated with Claude Code

[saadqbal]

Really careful PR — the fail-closed properties all hold. A verification that ran and FAILED has no TRACEBLOC_ALLOW_UNVERIFIED escape (verify-blob runs outside the try/catch, $LASTEXITCODE gates the exit); the opt-out covers "cannot verify" only; cosign-won't-run is split from bad-signature via Test-CosignRuns and gives ARM-aware advice; and -eq '1' avoids the [bool]'0' foot-gun. release.yml already signs both Windows arches and publishes .sig/.cert where install.ps1 looks (identity/issuer match install.sh), so mandatory-verify doesn't break existing installs. Ran the harness — one non-blocking test-hermeticity nit inline; CI is green.