Fail closed on unreadable scala-cli Bloop evidence, as sbt does (#1270)

[Mikola Lysenko (mikolalysenko)]

LLM Description written by Claude Code:claude-opus-5-5

Fixes #1270

Summary

crawlers::scala_evidence::discover now fails closed when a Bloop project file can't be read whole. That covers a file over MAX_FILE_BYTES, a read error, and a truncated or non-Bloop record. It now uses the policy that formats::sbt::evidence::resolution and crawlers::sbt_evidence already use. The skip-and-continue branches are deleted.

Why

• Issue #1270, register row E95 (register/10-audit-ecosystems.md), living document Part 5 (vendored JVM gates).
• The two JVM evidence readers had one policy each for evidence they couldn't read. sbt said "no evidence at all, never a partial resolution". scala-cli skipped the bad file and built a resolution from whatever files remained. When the skipped file was the newest project's -test twin, coursier_gate::gate judged the main project alone, so a //> using test.dep at another version passed as ok.
• Leverage: B 1, U 0, D ≈0.5 (one fail-closed rule for both readers), R L, score 3. This was the top candidate whose files no open PR touches. The larger refactors (An unparseable .socket/manifest.json is reported under five different --json error codes depending on the command #931 cluster, Upstream restore sends every public-registry lookup at once; the registry_concurrency cap has no caller #1220, Delete the vendored and hosted-vlt helpers left without a production caller by the v5 consolidation #782, Hash agent-mode jar members through the shared streaming zip comparator instead of buffering each member #914, NuGet version identity is normalized by vendor but not by PurlKey, so a freshly vendored 4-part version is judged unused and pruned #1202) are still held by open PRs.

What changed

• discover: an oversized, unreadable or non-Bloop .json returns None. Non-regular entries (symlinks, FIFOs), non-.json names and other workspaces' projects are still ignored.
• Module and MAX_FILE_BYTES docs now state the fail-closed rule.

Lines

Production +23/−6 (one file, about half of it doc comments). Tests +87/−1.

Behavior

• An unreadable project file anywhere in .scala-build/.bloop now makes the scala-cli gate skip with vendor_scala_cli_resolution_missing, where it used to pass or judge partial evidence. Already-vendored GAVs still re-plan, because the gate's vendored path doesn't need evidence.
• No change to codes, the contract, JSON shape or lockfile bytes.

Test evidence

• Red → green: crawlers::scala_evidence::tests::an_unreadable_project_file_hides_no_conflict covers the issue's table. With both files intact the gate reports vendor_scala_cli_version_conflict. A truncated, foreign or oversized -test twin, or a truncated newest main project, makes discover return None and the gate report vendor_scala_cli_resolution_missing.
  • It fails on main's discover: Some(ScalaEvidence { files: ["…/sc_p.json"] … }) instead of None.
  • It passes on this branch.
• caps_and_special_files now also asserts that an oversized file next to a readable project gives None, and that removing it brings the evidence back.
• cargo clippy --workspace --all-features -- -D warnings: clean.
• cargo test -p socket-patch-core --lib: 6092 passed. The 4 failures are the known root-sandbox ones that also fail on main (relax_loop_must_not_traverse_symlinked_root, an_unremovable_hidden_lock_keeps_every_store_entry, wire_write_failure_maps_error_and_leaves_lock_untouched, wire_failure_rolls_back_already_written_files).
• cargo test -p socket-patch-cli --all-features --test e2e_scala_cli_vendor: 31 passed, 1 ignored. --test spawn_env_hygiene: 12 passed.

Risk

Low. The change is one function, and it only adds refusals in a case where the result used to be partial evidence.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YTFbfEEtE5pUm5pdEaFyMz

---

Note

Low Risk
Single-function policy tightening in evidence discovery only increases refusals when Bloop metadata is incomplete; no auth, wire format, or gate code contract changes.

Overview
Fixes #1270 by changing scala-cli Bloop evidence discovery to fail closed when any .scala-build/.bloop project JSON cannot be read whole—over MAX_FILE_BYTES, I/O failure, truncated content, or non-Bloop JSON—instead of skipping that file and returning partial resolution evidence.

That matches the sbt evidence reader: an unreadable record might be the one that names a conflicting GAV (e.g. a -test twin with test.dep at another version), so the vendor gate must not judge main-only evidence. Unreadable cases now yield no evidence (discover → None, gate → vendor_scala_cli_resolution_missing) rather than incorrectly passing when the bad file hid a version conflict.

Docs for the cap behavior are updated; tests cover oversized files beside valid projects and gate behavior for truncated/foreign/oversized -test twins and corrupt newest main projects.

^{Reviewed by Cursor Bugbot for commit d762714. Configure here.}

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

BugBot review

---

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 d762714. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Ready for review at d76271406fc5.

• CI: required checks ci-ok and clippy green; 7 check suites succeeded, 2 skipped. 2 superseded workflow run(s) show as cancelled; the required gates passed on this head.
• Mergeable against main, no CHANGELOG.md change.
• Bugbot reviewed this head; no unresolved review threads.

Labeled Ready for review by the burn-down agent. Slack announcement pending (connector unavailable this run).

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

[agent] ci-ok failed on a5fbe0c (the main merge) only because its upstream jobs were cancelled: 12 jobs, including e2e, cargo, coverage, hosted-e2e and node-addon, were cancelled within about 3 minutes of the run starting, and none of them failed a test. The previous head d762714 was green on all 100 checks. I re-ran the failed jobs of run 37994113413 once. If they fail again, that failure is real and I'll take it up here.

---

Generated by Claude Code