Decision: Select a safe treatment for glib GHSA-wrw7-89jp-8q8g

[Niko4417]

Lifecycle contract: docs/qa/issue-lifecycle.md.

Canonical lifecycle states: status: new, status: triaged, status: ready,
status: in progress, status: pr open, status: ready for human review, status: blocked,
status: waiting for user, and status: done.

Parent Epic: None — standalone supply-chain security decision; tracked alongside #49 without becoming a lifecycle-activation child

Planning contract

• Contract version: v1
• Change classification: mandatory-delta

Implementation Ready freezes the decision question, authority, constraints, options, evaluation
method, acceptance criteria, evidence requirements, and delivery target for this version. A semantic
change increments the version and returns the issue to status: new until readiness validation
succeeds again. A planning agent may request validation directly after an evidence-first grilling
session has resolved the product inputs; no separate GitHub approval or human comment is required.

Decision question and authority

• Decision to resolve: Select and prove the smallest safe, supportable treatment for Dependabot alert
  ci: establish Keiko-parity quality gates #1 / GHSA-wrw7-89jp-8q8g affecting locked glib 0.18.5, without weakening dependency gates,
  hiding the alert, expanding Linux into the supported platform set, or prescribing an unproven
  dependency migration.
• Why it must be decided now: The repository has an open medium-severity alert over its authoritative
  cross-target Cargo lock. The current declared deliverable is macOS arm64, but the vulnerable crate
  remains in the deferred Unix/GTK graph and no routine lock update can cross the GTK 0.18 constraint.
• Decision owner: Keiko Native maintainers, with security and release evidence.
• Agent Planning Baseline, Decision Addendum, parity, risk, or incident reference:
  Agent Planning Baseline security/privacy, verification/release evidence, and supply-chain/licence
  drift risk; Decision Addendum supported-platform boundary; Dependabot alert ci: establish Keiko-parity quality gates #1.
• Relevant accepted ADRs and repository evidence: ADR-0001; ADR-0006; protected dev commit
  f65ee963a66c1c92bda60f3dacd7a2f1f5a603ba;
  Dependabot alert #1;
  GHSA-wrw7-89jp-8q8g;
  RUSTSEC-2024-0429.
• Required resulting record: security decision and ADR-0013; separately accepted implementation issue for any repository or dependency mutation

Current evidence and uncertainty

• native/Cargo.lock at the baseline resolves glib 0.18.5.
• The advisory affects glib >=0.15.0,<0.20.0; the first patched release is 0.20.0. The defect
  is unsoundness in VariantStrIter iterator implementations and can result in undefined behavior
  and null-pointer crashes in optimized builds.
• cargo tree --locked --offline --target all -i glib@0.18.5 reaches the package through the
  Tauri 2.11.5 / Tauri runtime / Wry 0.55.1 GTK 0.18 and WebKitGTK graph.
• The same inverse query for aarch64-apple-darwin and x86_64-pc-windows-msvc prints no
  dependency path. This proves target-graph absence for the inspected baseline, not permanent
  safety or authority to ignore the cross-target lock.
• quality/project.json currently declares only the macOS arm64 productive target. Windows remains
  a first-class product platform but is not yet a declared productive target; Linux remains
  explicitly deferred.
• Tauri #12048 was closed as not planned.
  Tauri #12564 records that Tauri will not maintain
  a GTK3 fork and tracks GTK4 migration for v3.
• gtk3-rs is archived. The gtk-rs maintainer states that
  0.18 is EOL and will receive no new release.
• No released, supportable Keiko dependency change has yet been proven. A direct glib lock bump,
  alert dismissal, ignore entry, warning-mode conversion, universal-lock scan removal, or threshold
  reduction is not authorized by this issue.

Constraints and non-negotiables

• Product and user constraints: Preserve a secure, reproducible desktop product on the declared
  macOS target and the approved future Windows boundary; do not add Linux acceptance or runtime
  support.
• Architecture and dependency constraints: Keep the Tauri host decision and current supported
  runtime behavior unchanged unless equal compatibility evidence justifies an ADR amendment.
  Preserve one authoritative, reproducible dependency model and do not create an unmanaged fork or
  parallel hidden lock graph.
• Security, privacy, regulatory, and operational constraints: The alert and advisory remain visible
  throughout evaluation. Do not dismiss, suppress, ignore, downgrade, or weaken Dependabot, OSV,
  Dependency Review, Socket, SBOM, provenance, signing, or release gates. Treat optimized-build
  undefined behavior as security-relevant even when current supported target graphs exclude it.
• Supported platforms and reference environments: Current declared aarch64-apple-darwin;
  compatibility check for the approved future x86_64-pc-windows-msvc boundary; --target all
  only as inventory evidence; Linux is deferred and not acceptance authority.
• Explicit non-goals: Implementing a dependency change; modifying PR docs: govern authenticated lifecycle overflow recovery #171; modifying the Codex
  lifecycle-tracer task; activating Linux; dismissing Dependabot alert ci: establish Keiko-parity quality gates #1; adding an advisory ignore;
  selecting Tauri v3, GTK4, a fork, vendoring, or manifest restructuring without equal evidence.

Evaluation journey

• Applicability: Not applicable — security and supply-chain decision with no user-facing UI change
• Actor and user goal: A security/release maintainer chooses a treatment that keeps shipped
  macOS/Windows composition safe and dependency evidence truthful.
• Representative starting state and sanitized test data: Protected dev baseline above, committed
  manifests and locks, public advisories, and public upstream metadata only.
• Observable path and failure or recovery states: Reproduce the all-target and target-specific
  graphs; evaluate every option; reject unavailable, unmaintained, incompatible, unverifiable, or
  gate-weakening treatments; preserve the last accepted dependency baseline until a separate
  implementation contract is ready.
• Platform variants: macOS arm64 is authoritative today; Windows compatibility must not be
  foreclosed; Linux results are inventory/diagnosis only.
• Evidence produced consistently for every affected option: Exact versions and target graphs,
  released-version and maintenance status, provenance/licence/SBOM impact, supported-target build
  and test feasibility, quality-gate impact, migration/removal cost, failure modes, and residual
  risk.

Options

Option	Description	Expected benefit	Cost or risk	Rejection condition
A — released supported upgrade	Move only to a released Tauri/Wry/GTK composition whose resolved supported-target graph removes the vulnerable range	Maintained upstream path and ordinary dependency provenance	May require breaking host/runtime migration; no proven release currently exists	Reject if unreleased, unsupported, Linux-expanding, behavior-changing without evidence, or still resolves affected glib
B — target-accurate graph removal	Restructure manifests/features/locks so the deferred GTK3 graph is absent from authoritative shipped composition while full dependency inventory remains truthful	Removes non-shipped vulnerable bytes without adopting an upstream fork	Cargo/Tauri constraints may make the split complex or create parallel dependency truth	Reject if any supported artifact can reach the graph, the alert is merely hidden, or scanning/SBOM/provenance weakens
C — maintained patched source	Adopt a verifiable maintained patch/fork only through a separately owned dependency contract	Could fix the exact code without waiting for Tauri v3	High provenance, maintenance, licence, update, and incident-response burden	Reject without durable ownership, immutable source identity, upstream-diff audit, supported-platform tests, SBOM/signing evidence, and removal plan
D — visible bounded deferral	Retain the current cross-target lock and open alert while proving supported artifacts cannot reach it and defining exact re-evaluation triggers	Smallest immediate product change and preserves alert visibility	Leaves known vulnerable/EOL bytes in the repository resolve graph and may become stale as targets change	Reject if any declared or packaged target reaches it, policy requires repository-wide removal, proof is incomplete, or triggers/owner/expiry are absent

No suppression, alert dismissal, warning mode, lowered threshold, or advisory exception is an option.

Evaluation plan

• Method: research + threat-model evaluation + compatibility test + operational evaluation
• Time or scope box: One decision issue; no productive dependency mutation. Stop after a reproducible
  recommendation and accepted ADR/security record.
• Workloads, scenarios, or attack paths: All-target graph inventory; current macOS graph; approved
  future Windows graph; optimized VariantStrIter reachability; package/SBOM inclusion; upgrade,
  removal, fork, and deferred-state failure/recovery.
• Metrics, thresholds, and measurement method: Zero affected glib in any accepted shipped
  composition; zero gate weakening or hidden advisory; reproducible locked resolution; maintained
  provenance; no Linux scope expansion; complete supported-target quality feasibility.
• Reference hardware, operating systems, versions, and dependencies: Protected baseline commit;
  Node.js 24.18.x/npm 11.16.x; repository Rust toolchain; Tauri 2.11.5; Wry 0.55.1; GTK 0.18.2;
  glib 0.18.5; macOS arm64; Windows target graph compatibility.
• Required primary sources and repository evidence: GitHub alert/advisory, RustSec advisory and fix,
  upstream release manifests and maintenance statements, committed Cargo manifests/lock,
  quality/project.json, target-aware dependency workflow, and retained command output with no
  private data.
• Reproduction commands and retained artifacts:

  cargo tree --locked --offline --manifest-path native/Cargo.toml --target all -i glib@0.18.5
  cargo tree --locked --offline --manifest-path native/Cargo.toml --target aarch64-apple-darwin -i glib@0.18.5
  cargo tree --locked --offline --manifest-path native/Cargo.toml --target x86_64-pc-windows-msvc -i glib@0.18.5
  npm run native:dependencies
  npm run native:security
  npm run quality
  npm audit --audit-level=high
  

• Testability, automation harness, CI ownership, and production-artifact isolation: Use detached
  temporary worktrees and repository-owned dependency/security commands. Experimental manifests or
  patches remain outside productive roots, are never pushed, and are deleted after evidence capture.
  Any selected mutation receives a separate accepted issue and exact-head CI.
• Bias, uncertainty, and evidence limitations: Dependabot evaluates the cross-target lock while the
  repository's OSV gate intentionally evaluates the declared target graph. Target absence does not
  prove future absence, non-reachability of every code path, or acceptability of EOL dependencies.
  Upstream plans and unreleased branches are not implementation evidence.

Experimental code is evidence only. It may exist only in a private temporary detached worktree and
must be deleted before this decision completes. Nothing becomes production architecture without a
separately accepted implementation issue.

Execution Authority

The repository-wide defaults in AGENTS.md apply and are not repeated here.

• Authorized repository: oscharko-dev/Keiko-Native
• Exact delivery target: dev (standalone)
• Allowed write scope: docs/adr/ADR-0013-glib-advisory-treatment.md,
  docs/qa/quality-gates.md, and quality/contract.test.mjs only when required to record and pin
  the selected decision; temporary evaluation artifacts may exist only outside the repository
• Additional prohibited paths or actions: All manifests, lockfiles, workflows, productive source,
  PR docs: govern authenticated lifecycle overflow recovery #171, Codex lifecycle-tracer work, repository settings, branch protection, Dependabot state,
  advisory exceptions, and dependency versions
• Authorized external mutations: Create/update this issue and a decision pull request only; read
  public upstream evidence and Dependabot alert ci: establish Keiko-parity quality gates #1; no alert dismissal, upstream mutation, release,
  label-policy, or repository-setting mutation
• Required credentials or secrets: Existing GitHub repository access only; no new secret or
  dependency-registry credential
• Delivery authority: human-only manual merge to dev
• Additional stop or escalation conditions: Stop if a supported remediation cannot be reproduced;
  if evaluation requires Linux product scope, a new identity/secret, an unmanaged fork, gate
  weakening, alert dismissal, or an implementation mutation; or if another issue/PR begins owning
  the same advisory treatment

The executing agent chooses a dedicated source branch using its own runner prefix. It must include
the issue number, remain unique to this issue, and be recorded in the pull request.

Decision matrix

Weights are fixed before option evaluation and must not change after results are known without
incrementing the contract version and returning to status: new.

Criterion	Weight	Option A	Option B	Option C	Option D	Evidence
Supported-artifact security correctness	30	TBD	TBD	TBD	TBD	Target graphs, optimized-path analysis, packaged inventory
Gate and alert integrity	20	TBD	TBD	TBD	TBD	Dependabot/OSV/Dependency Review/Socket contract
macOS and Windows compatibility	20	TBD	TBD	TBD	TBD	Locked builds, tests, packaging feasibility
Provenance and maintenance ownership	15	TBD	TBD	TBD	TBD	Releases, signatures, licences, SBOM, upstream status
Reproducibility and operational clarity	10	TBD	TBD	TBD	TBD	Hermetic commands, update/recovery/runbook evidence
Scope and lifecycle cost	5	TBD	TBD	TBD	TBD	Migration, removal, support, and follow-up estimate

Acceptance criteria

• AC1 — Every viable option is evaluated against the fixed matrix and the same primary-source,
  target-graph, supported-platform, provenance, and quality-gate evidence bar.
• AC2 — The recommendation follows reproducibly from recorded evidence and distinguishes current
  target absence from future platform safety.
• AC3 — ADR-0013/security decision records the selected treatment, rejection rationale,
  residual risk, owner, review triggers, and separately scoped implementation issue(s).
• AC4 — No dependency, lock, workflow, alert, suppression, exception, threshold, branch
  protection, PR docs: govern authenticated lifecycle overflow recovery #171, or lifecycle-tracer mutation occurs under this decision authority;
  temporary artifacts are deleted.

Verification commands

cargo tree --locked --offline --manifest-path native/Cargo.toml --target all -i glib@0.18.5
cargo tree --locked --offline --manifest-path native/Cargo.toml --target aarch64-apple-darwin -i glib@0.18.5
cargo tree --locked --offline --manifest-path native/Cargo.toml --target x86_64-pc-windows-msvc -i glib@0.18.5
npm run native:dependencies
npm run native:security
npm run quality
npm audit --audit-level=high

Definition of Ready

• The decision question, owner, urgency, constraints, options, and non-goals are explicit.
• Evaluation methods, equal-workload criteria, thresholds, environments, and evidence sources
  are defined before results are collected.
• Execution Authority and experimental-code disposal are explicit.
• The planned evaluation is executable in the available environment.

Definition of Done

• Every criterion has attributable, reproducible evidence and recorded limitations.
• The recommendation, dissenting evidence, risks, and residual uncertainty are explicit.
• The resulting decision record is accepted through its normal governance path.
• No throwaway code, secret, private evidence, alert dismissal, or undeclared productive
  dependency remains.
• Documentation and follow-up implementation issues reflect the decision without expanding it.

Stop conditions

• Stop when the current contract version does not match its automated readiness record.
• Stop when any option, criterion, threshold, platform, boundary, or authority changes.
• Stop when representative evidence cannot be produced in the declared environment.
• Stop when evaluation would require customer data, an undeclared secret, an unauthorized external
  mutation, alert dismissal, advisory suppression, gate weakening, or Linux product scope.
• Stop before turning experimental evidence into productive architecture or feature delivery.