Patch 14 security advisories with conservative lockfile bumps

[minerva-sky]

What

Lockfile-only security bump. bundler-audit flags 14 advisories across five gems in Gemfile.lock:

Gem	Locked	Bumped	Advisories
faraday	2.11.0	2.14.3	SSRF via protocol-relative URLs (CVE-2026-25765 + incomplete-fix follow-up CVE-2026-33637), nested-params stack-exhaustion DoS (CVE-2026-54297)
uri	0.13.1	1.1.1	userinfo credential leakage (CVE-2025-27221) and its bypass (CVE-2025-61594)
concurrent-ruby	1.3.4	1.3.8	AtomicReference NaN livelock (CVE-2026-54904), RW-lock exclusivity bugs (CVE-2026-54905, CVE-2026-54906)
addressable	2.8.7	2.9.0	template ReDoS (CVE-2026-35611)
rexml	3.3.1	3.4.4	five DoS advisories (CVE-2024-39908 … CVE-2024-49761)

faraday matters most: it's ruby-openai's transport, i.e. the path every LLM call takes. uri and concurrent-ruby are runtime-transitive; addressable and rexml are dev-only (webmock/rubocop).

A sixth gem moves with no advisory behind it: faraday-multipart 1.0.4 → 1.2.0. Its 1.0.4 gemspec declares required_ruby_version < 4, so a lock holding it cannot bundle install on current stable Ruby (4.x) at all — the panel caught it riding along undisclosed, and the root cause turned out to be exactly that ceiling. 1.2.0 lifts it and keeps the Faraday::UploadIO alias that ruby-openai 7.1.0's multipart path calls (verified present in 1.2.0's source). WORLD.md: "Track the latest stable Ruby."

How

bundle lock --update addressable rexml concurrent-ruby uri faraday faraday-multipart --conservative — all six bumped in place. ruby-openai stays at 7.1.0; the gemspec is untouched and Gemfile.lock is excluded from the packaged gem, so consumers see nothing. This patches only the gem's own dev/CI resolution.

Verification

• bundler-audit check: No vulnerabilities found (was 14)
• bundle exec rake (610 examples + standardrb): exit 0
• Caveat, stated plainly: local verification ran on Ruby 4.0.6 (no 3.2 interpreter on this machine). CI's 3.2.4 run is the adjudicating evidence — it's where uri 1.1.1 activates as a gem over the 0.12.x default, the largest behavioral delta here. Please treat the green Ruby 3.2.4 check, not the local run, as the merge evidence.

One review finding didn't block but is tracked: faraday ≥ 2.12 embeds the full request URL in error messages, which agentic's logger and execution journal persist verbatim — a query-string-auth exposure for custom api_base_url gateways. Filed as item 3 on #12 with a proposed sanitize-at-boundary fix.

Panel

• DHH — cleared. "A Gemfile.lock-only conservative bump has no API surface to be un-Ruby-ish"; flagged the faraday-multipart drift and the uri major jump, judged both acceptable.
• Obie Fernandez — objected-then-cleared. Three objections: the undisclosed faraday-multipart bump, verification on the wrong Ruby, and faraday 2.14's URL-bearing error messages hitting persistent sinks. Second pass verified the fixes claim-by-claim (forced sixth bump disclosed, CI-as-evidence stated verbatim in the commit, no failure text reaches LLM prompts, exposure tracked on Harden PersistentAgentStore path handling and capability-filter dispatch against untrusted agent configs #12): "two faraday SSRF CVEs outrank a same-trust-domain local-log refinement."
• Vladimir Dementyev — objected-then-cleared. Two objections: the sixth bump and the 4.0.6-vs-3.2.4 evidence gap. Second pass independently confirmed faraday-multipart 1.0.4's < 4 ruby ceiling forces the move, and ran Faraday::UploadIO resolution live under 1.2.0 to close the seam VCR can't see. "Refusing to peg a shared host's CPU building a second interpreter for evidence CI produces for free is the right performance call."
• The gem consumer — cleared. Verified Gemfile.lock is excluded from the packaged gem, gemspec constraints untouched, version.rb untouched: "the diff moves the single tested point, it doesn't change the contract surface."