Cut E2E artifact size by 53% with shared compression

[Mikola Lysenko (mikolalysenko)]

Every E2E consumer currently downloads the same ZIP of test binaries. A recent merge-queue job spent 7m52s downloading its 200 MB bundle. Most recent runs took 6–7 seconds, so that delay is an outlier, but repeating the full transfer across roughly 150 consumers also amplifies artifact-service congestion.

Pack the same binaries into a Zstandard-compressed tar before upload. A shared 128 MiB compression window reuses common Rust code across executables, which per-file ZIP compression cannot do. Both E2E and cargo-vex consumers unpack it before their existing test steps; the test selection, build flags and executable bytes stay the same. The upload retry action accepts a compression level, defaulting to its existing level 6; these already-compressed bundles use level 0.

Measured with the actual 36-file Linux artifact and the exact workflow commands:

• Artifact size: 199,637,431 → 94,339,590 bytes (53% smaller).
• Local packing: 1.8 seconds; unpacking: about 0.4 seconds. The first hosted Linux producer produced a 94,272,299-byte artifact: seven seconds to compress and three to upload, versus 17–20 seconds for the previous ZIP upload step in recent runs. The Windows producer also passed (five seconds to compress, three to upload).
• Every filename and SHA256 matched after both consumer commands restored the bundle.
• All 149 hosted E2E consumers passed: 130 Linux and 19 Windows. Median download-plus-unpack time was three seconds on Linux and two seconds on Windows.

The first run also hit the existing Windows global-Gemfile fixture flake: gem env can time out before finding any packages. The branch includes the approved fixture fix from #1169, which has since merged into main. Both Windows unit-test shards now pass, with all refusal, byte-preservation and attestation assertions preserved.

Validation: all 275 Python harness tests pass (one skipped), including actual archive round trips, executable permissions, missing-input failure and corrupt-archive failure. Merging locally with #1143 and #1166 passes all 286 harness tests. The corrected Gemfile-refusal test passes both normally and with PATH=/nonexistent, so it no longer needs a working host gem command. Actionlint has the same 97 existing findings and no new ones; git diff --check passes.

The runner-provided tar and zstd tools avoid an additional installation step; Zstandard is listed in the Ubuntu, Windows and macOS images. The compression window adds bounded memory use during packing and unpacking. This complements #1143's independent platform scheduling and #1166's release-test sharding. The full hosted CI run passed on 77870a6c in 30m30s. Its final job was the unsharded release suite (29m06s); #1166 addresses that remaining wait. The compression measurements above establish the artifact-size and transfer improvements.

---

Note

Medium Risk
CI artifact format changes affect every E2E and cargo-vex consumer; mitigated by round-trip tests and unchanged binary layout after unpack.

Overview
Shrinks E2E binary artifacts by packing target/e2e-bin into a single zstd-compressed tar (shared long window) before upload, instead of uploading the raw directory as a per-file ZIP. Consumers download to target/e2e-archive, unpack into target/e2e-bin, then run the same staging steps as before.

The reusable upload-artifact action gains a compression-level input (default 6); the E2E bundle upload sets 0 so GitHub’s ZIP wrapper does not re-compress the .tar.zst.

Adds scripts/tests/test_ci_e2e_archive.py to exercise the real CI compress/unpack scripts (permissions, missing bundle, corrupt archive).

Stabilizes gem_hosted_global_gemfile_setting_is_refused by materializing an installed gem in the project so the refusal path does not depend on a flaky host gem env probe on Windows.

^{Reviewed by Cursor Bugbot for commit 77870a6. Configure here.}

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[final reviewer] Tanmay Singla (@Tanmay182003) One non-merge commit landed after your approval on 27ca53bb: 77870a6c Make gem global-gemfile refusal test hermetic (test-only). I've disabled auto-merge until you take another look at head 77870a6c6b.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

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 77870a6. Configure here.}

[Mikola Lysenko (mikolalysenko)]

[burn-down agent] Ready for review at 77870a6c6b

• CI: all 372 check runs on the current head green or skipped (ci-ok success).
• Mergeable against main, no conflicts; no CHANGELOG.md change.
• Cursor Bugbot reviewed 77870a6c (23:14 UTC): no issues found; no open review threads.
• Note for reviewers: Tanmay Singla (@Tanmay182003) approved 27ca53bb; one commit (77870a6c, the Windows gem fixture fix from Fix Windows flake in gem global-gemfile refusal e2e #1169, now on main) landed after that approval.

---

Generated by Claude Code