Vox Language Rules — Phase 2: Lint Extension with Stable Diagnostic IDs (2026-05-09)

Phase 2 — `vox-code-audit` Lint Extension

Parent plan: vox-language-rules-and-enforcement-plan-2026.md Depends on: Phase 1 Tasks 1–2 (diagnostic-catalog scaffolding + #[vox_diagnostic] macro). For agentic workers: REQUIRED SUB-SKILL: Use superpowers:executing-plans.

Goal: Add 14 new detectors to vox-code-audit, each with a stable ID, a serializable autofix descriptor, an --explain page, and a negative-example corpus entry. Add vox check --for-llm JSON output mode optimized for LLM agents quoting reproducers and applying fixes. Establish the diagnostic-shape contract that Phases 3–5 reuse.

Architecture: Detectors live in crates/vox-code-audit/src/detectors/<category>/<id>.rs. Each detector implements a Detector trait that returns Vec<Diagnostic> from a &CompiledModule input. Diagnostics carry the full LLM-friendly shape from the parent plan §“LLM-friendly diagnostic shape.” Autofix descriptors are serializable LintFix enums (no closures), so a CI log entry can be replayed by an LLM agent or a vox check --apply-fix <id> invocation.

The detector trait:

pub trait Detector {
    fn id(&self) -> DiagnosticId;
    fn check(&self, module: &CompiledModule, ctx: &DetectorCtx) -> Vec<Diagnostic>;
}

pub struct Diagnostic {
    pub id: DiagnosticId,
    pub severity: Severity,
    pub span: Span,
    pub excerpt: Excerpt,           // 3 lines context + offending line + 3 lines after
    pub message: String,            // human-readable
    pub rationale: &'static str,    // constant per ID; pulled from the catalog
    pub suggested_fix: Option<LintFix>,
    pub confidence: Confidence,     // Certain / Likely / Speculative
    pub alternatives: Vec<LintFix>, // additional plausible fixes
    pub adr: Option<&'static str>,
}

pub enum LintFix {
    Replace { range: Span, new_text: String },
    InsertBefore { offset: usize, text: String },
    InsertAfter { offset: usize, text: String },
    RemoveDecorator { name: String, span: Span },
    RenameSymbol { old: String, new: String, sites: Vec<Span> },
    AddImport { path: String, after_line: usize },
    Composite(Vec<LintFix>),
}

pub enum Confidence { Certain, Likely, Speculative }

Out of scope for Phase 2:

Type-system-level rules (Phase 3 owns Id[T], named errors, syntax_version).
Runtime traps (Phase 4).
Effect-system enforcement (Phase 5).
Renaming or removing existing vox-code-audit detectors. Existing skeleton/* detectors stay as-is.

Verification setup

cargo test -p vox-code-audit --lib detectors:: — per-detector unit tests.
cargo test -p vox-code-audit --test golden — input-.vox/expected-diagnostic-JSON snapshot tests.
cargo run -p vox-cli -- check examples/golden/ --json — should emit zero diagnostics from new detectors against the curated golden corpus (after burn-down PRs).
cargo run -p vox-cli -- check --for-llm examples/golden/anti/ --json — should emit exactly the diagnostics declared in each anti-example’s expected_diagnostic: frontmatter.
cargo run -p vox-cli -- check --explain vox/llm/direct-provider-call — should print the rationale and a worked-example fix.

Task 1: Detector trait + catalog populated with all 14 IDs

Files:

Modify: crates/vox-code-audit/src/diagnostics/catalog.rs — add 14 variants, each with #[vox_diagnostic(...)].
Modify: crates/vox-code-audit/src/detectors/mod.rs — declare 14 detector submodules (skeletons returning empty Vec).
Create: docs/src/reference/diagnostics/index.md (hand-written index)
Create (one each): docs/src/reference/diagnostics/<id>.md (skeleton with frontmatter + headings)

Diagnostic IDs and metadata (the full catalog landed in this task; implementations land in Tasks 2–14):

ID	Severity at land	Severity at escalation	Phase	ADR	Audit item
`vox/llm/direct-provider-call`	error	(already error)	2	TBD-llm-call-discipline	A.5
`vox/secret/env-get-shape`	error	(already error)	2	TBD-secret-discipline	A.6
`vox/crypto/banned-crate-import`	error	(already error)	2	cryptography-ssot-2026	A.30
`vox/style/duplicate-prefix-name`	warning	warning (perma)	2	—	A.16
`vox/style/long-range-coupling`	note	warning (next minor)	2	—	A.17
`vox/control-flow/question-mark-opportunity`	note	warning (next minor)	2	—	A.19
`vox/control-flow/option-combinator-opportunity`	note	warning (next minor)	2	—	A.20
`vox/doc/missing-adr-citation`	note	warning in `vox-actor-runtime`/`vox-orchestrator`/`vox-compiler`	2	—	A.22
`vox/decorator/position-mismatch`	warning	error (next minor)	2	—	A.27
`vox/require/justification-prose-required`	note	warning (next minor)	2	—	A.28
`vox/handler/panicking-builtin`	warning	error (after 2 minors)	2	ADR-019	A.29 (descoped from A.4)
`vox/state-machine/unreachable-state`	warning	warning (perma — proof-tier deferred)	2	ADR-030	A.54 (descoped)
`vox/secret/leaked-to-span`	error	(already error)	2	telemetry-trust-ssot	A.11 (descoped from full taint)
`vox/auth/endpoint-missing-decorator`	warning	error after Phase 5 effect system lands	2	TBD-endpoint-auth	A.58 (descoped)

Why these and only these: Each is enforceable with the current compiler surface (no Phase 3/5 dependency) and addresses a known LLM-confusion or security smell. The detectors deferred to other phases are listed in §“Out of scope” of the parent plan.

Verify: cargo build -p vox-code-audit passes; cargo run -p vox-cli -- check --explain <any-id-above> returns the rationale from the catalog (skeleton text initially; populated by Tasks 2–14).

Task 2: `vox/llm/direct-provider-call` detector

Files:

Modify: crates/vox-code-audit/src/detectors/llm/direct_provider_call.rs
Create: examples/golden/anti/llm-direct-provider-call.vox
Modify: docs/src/reference/diagnostics/llm-direct-provider-call.md — fill in rationale prose

Suggested fix (Confidence::Likely):

Replace { range: <call site>, new_text: "populi.complete(...)" }

with alternatives listing other populi.* shapes (populi.stream, populi.embed).

Negative-example file:

// vox:skip examples/golden/anti/llm-direct-provider-call.vox
// frontmatter:
// ---
// error_kind: lint
// expected_diagnostic: "vox/llm/direct-provider-call"
// mens_role: anti-example
// ---

fn ask_claude(prompt: str) -> Result[str, NetError] {
    std.http.post_json(
        url: "https://api.anthropic.com/v1/messages",
        body: { model: "claude-opus-4-7", input: prompt },
    )
}

Why this is error immediately: Direct provider calls bypass vox-populi’s telemetry, retry, cost accounting, and capability ledger. This is a security and observability surface, not a style preference.

Suppression: // toestub-ignore(vox/llm/direct-provider-call) — <reason> on the line above the call. Reviewer audit required.

Verify: Golden test in crates/vox-code-audit/tests/golden/llm-direct-provider-call/; --explain returns the rationale.

Task 3: `vox/secret/env-get-shape` detector

Files:

Create: crates/vox-code-audit/src/detectors/secret/env_get_shape.rs
Create: examples/golden/anti/secret-env-get-shape.vox
Modify: docs/src/reference/diagnostics/secret-env-get-shape.md

Suggested fix (Confidence::Certain):

Replace {
  range: <env.get call>,
  new_text: "vox_secrets.resolve(\"<derived-secret-id>\")"
}

where <derived-secret-id> is computed by lowercasing + snake-casing the env var name. The fix is Likely (not Certain) when a matching SecretId does not yet exist in crates/vox-secrets/src/spec.rs — the diagnostic carries an alternatives entry suggesting addition of a SecretSpec.

Composite fix shape when no SecretId exists:

Composite([
    InsertBefore { offset: <spec.rs first var line>, text: "    SecretSpec::new(SecretId::<NewId>, ...),\n" },
    Replace { range: <env.get>, new_text: "vox_secrets.resolve(SecretId::<NewId>)" }
])

LLM-target note: This is exactly the kind of multi-file fix that benefits from the Composite LintFix shape. An LLM agent can see “I need to update spec.rs and the call site” without inferring it from prose.

Verify: Golden tests for: (a) literal matching pattern, (b) literal not matching, (c) non-literal arg (no fire), (d) suppression with reason works.

Task 4: `vox/crypto/banned-crate-import` detector

Files:

Create: crates/vox-code-audit/src/detectors/crypto/banned_crate_import.rs
Modify: deny.toml — add [[bans]] entries for aegis, ring, transitive cmake/nasm-pulling crates (audit item [A.30])
Create: examples/golden/anti/crypto-banned-crate.vox

Rule: This is dual-enforced: deny.toml for Rust crate dependencies (cargo-deny), and the vox-code-audit detector for import statements in .vox files referencing banned cryptography surfaces. The Vox-side detector matters because future .vox plugins may pull in Rust crates indirectly via @host_crate(...) style imports.

Why dual enforcement: [AGENTS.md:95–98] is currently prose; deny.toml enforces the Rust path; this detector enforces the Vox path. Symmetric coverage.

Verify: Synthetic .vox file importing a banned name fires; deleting the import passes; cargo deny check integrates into the same CI job.

Task 5: `vox check --for-llm` JSON mode

Files:

Modify: crates/vox-cli/src/check.rs (or appropriate command file) — add --for-llm flag
Modify: crates/vox-code-audit/src/report.rs — add LlmReport struct alongside the existing JsonReport
Create: crates/vox-code-audit/tests/llm_mode_smoke.rs

Why this is the largest single LLM-target delta: A typical compiler emits a diagnostic and assumes the human reader has the file open. An LLM agent invoked from a CI log or a tool-call response often does not have the file. The --for-llm mode includes everything the agent needs to propose a fix without further context fetches:

{
  "schema": "vox.lint.llm-report.v1",
  "vox_version": "0.6.0",
  "diagnostics": [
    {
      "id": "vox/llm/direct-provider-call",
      "severity": "error",
      "since": "0.6.0",
      "adr": "TBD-llm-call-discipline",
      "file": "src/agents/researcher.vox",
      "line": 42,
      "column": 8,
      "message": "Direct provider call to api.anthropic.com bypasses vox-populi",
      "rationale": "Direct HTTP calls to LLM provider hostnames bypass the populi telemetry, cost accounting, capability ledger, and retry policy. All inference traffic must route through populi.*.",
      "minimal_repro": "fn f() {\n    std.http.post_json(url: \"https://api.anthropic.com/v1/messages\", body: {})\n}",
      "excerpt": {
        "lines": [40, 41, 42, 43, 44],
        "text": "fn ask_claude(prompt: str) -> Result[str, NetError] {\n    std.http.post_json(\n        url: \"https://api.anthropic.com/v1/messages\",\n        body: { model: \"claude-opus-4-7\", input: prompt },\n    )\n}"
      },
      "suggested_fix": {
        "kind": "replace",
        "range": { "start": [41, 4], "end": [44, 5] },
        "new_text": "populi.complete(model: \"claude-opus-4-7\", prompt: prompt)"
      },
      "confidence": "likely",
      "alternatives": [
        { "kind": "replace", "new_text": "populi.stream(...)" },
        { "kind": "replace", "new_text": "populi.embed(...)" }
      ],
      "explain_url": "https://vox-lang.org/diag/vox/llm/direct-provider-call"
    }
  ]
}

Compared with existing --json mode: --for-llm adds minimal_repro, excerpt.text, rationale, confidence, alternatives, and explain_url. The existing --json shape is preserved as-is for tools that already consume it.

Verify: Snapshot test of the JSON output for one diagnostic per ID in the catalog (after Tasks 2–14 land their detectors).

Task 6: `vox check --explain <id>` subcommand

Files:

Create: crates/vox-cli/src/explain.rs
Modify: crates/vox-cli/src/main.rs — wire the subcommand
Modify: docs/src/reference/diagnostics/<id>.md files — content authored in Tasks 2–14

Output shape:

$ vox check --explain vox/llm/direct-provider-call

vox/llm/direct-provider-call (error, since 0.6.0)
ADR: TBD-llm-call-discipline
URL: https://vox-lang.org/diag/vox/llm/direct-provider-call

Why this rule exists
--------------------
Direct HTTP calls to LLM provider hostnames bypass the populi telemetry,
cost accounting, capability ledger, and retry policy. All inference traffic
must route through populi.*.

Bad
---
fn ask_claude(prompt: str) -> Result[str, NetError] {
    std.http.post_json(
        url: "https://api.anthropic.com/v1/messages",
        body: { model: "claude-opus-4-7", input: prompt },
    )
}

Good
----
fn ask_claude(prompt: str) -> Result[str, NetError] {
    populi.complete(model: "claude-opus-4-7", prompt: prompt)
}

Suppress (audit-required)
-------------------------
// toestub-ignore(vox/llm/direct-provider-call) — <reason>

Why bad/good is structured prominently: LLMs asked “fix this Vox diagnostic” do dramatically better with a worked example pair than with prose alone. This format is the same input shape an LLM would naturally produce for in-context learning.

Verify: vox check --explain succeeds for every ID in the catalog; CI fails if any catalog ID lacks an --explain page.

Task 7: `vox/doc/missing-adr-citation` detector

Files:

Create: crates/vox-code-audit/src/detectors/doc/missing_adr_citation.rs

Rule: For every public fn, actor, workflow, activity, @table, @endpoint in scope of crates/vox-actor-runtime/**, crates/vox-orchestrator/**, crates/vox-compiler/**: the /// doc must contain at least one match of (ADR-\d+|TASK-\d+\.\d+|Phase \d+). Outside those scopes: same rule fires as note, escalates to warning per minor.

Anti-T-number rule: Active rejection of T\d+ references in /// doc as a corpus-drift signal. Honors anti-recommendation [A.69].

Suggested fix: Cannot autofix (the human must pick the citation). Diagnostic emits a list of recently-modified ADRs and TASKs touched in the same PR as candidates (computed from git log of the file).

Why scope matters: vox-actor-runtime/vox-orchestrator/vox-compiler are the highest-stakes correctness surfaces; doc drift there is more dangerous than in CLIs or examples.

Verify: Golden test with: (a) properly-cited fn, (b) uncited fn fires, (c) T-number cited fires the anti-rule.

Task 8: `vox/decorator/position-mismatch` detector

Files:

Create: crates/vox-code-audit/src/detectors/decorator/position_mismatch.rs

Rule: Detects:

A bare keyword used in a position where a decorator is expected (durable fn instead of @durable fn).
A decorator used in a position the grammar would parse as a bare-keyword declaration.
A decorator that is redundant given the bare keyword (e.g., @actor actor MyActor).

Suggested fix (Confidence::Certain for simple cases):

For durable fn → Replace { range: <durable>, new_text: "@durable" } plus space normalization.
For redundant @actor actor X → RemoveDecorator { name: "@actor", span: <decorator span> }.

LLM-target note: Decorator-position errors are the most common Vox parse error LLMs produce. Surfacing them with a Certain-confidence autofix means an agent can fix the file in one tool call.

Verify: Golden tests covering all four mismatch shapes.

Task 9: `vox/style/duplicate-prefix-name` detector

Files:

Create: crates/vox-code-audit/src/detectors/style/duplicate_prefix_name.rs

Rule: Identifier matches (\w+)_\1(?:_|$) — e.g., user_user_id, tasks_tasks, task_task_status. Fires warning with Confidence::Likely autofix that strips the duplication (user_user_id → user_id). The fix is Likely not Certain because the duplication may be intentional (e.g., a join table column user_user_id referring to the FK from users.users); alternatives includes “no fix — keep as-is.”

Why it matters for LLM target: This is a high-signal LLM-confusion smell. Models that hallucinate column names often produce duplicated prefixes; flagging early kills the smell at the lint layer.

Verify: Golden tests for matching/non-matching identifiers; rename fix updates all callsites within the module via RenameSymbol.

Task 10: `vox/style/long-range-coupling` detector

Files:

Create: crates/vox-code-audit/src/detectors/style/long_range_coupling.rs

Rule: Warn when:

A let-binding’s definition span end and its last use span start are > 80 lines apart.
A name shadows another in a nested scope at any nesting level > 2.

Suggested fix: None (refactoring is too contextual). Diagnostic includes excerpt of both the definition and the farthest use.

Why it matters: AI generators frequently produce long single functions where local state is defined far from its use, confusing later edits. P3 (LANGUAGE_DESIGN_PRIORITIES.md) made checkable.

Verify: Golden test with synthetic 100-line function that uses a binding from line 3 at line 95.

Task 11: `vox/control-flow/question-mark-opportunity` + `option-combinator-opportunity`

Files:

Create: crates/vox-code-audit/src/detectors/control_flow/question_mark_opportunity.rs
Create: crates/vox-code-audit/src/detectors/control_flow/option_combinator_opportunity.rs

Rules:

Question-mark: match foo() { Ok(x) => x, Err(e) => return Err(e) } → foo()?. Mirror of clippy::question_mark. Autofixable with Confidence::Certain.
Option combinator: match opt { Some(x) => f(x), None => default } → opt.map(f).unwrap_or(default). Autofix Confidence::Likely (because side-effects in f may matter).

Why both at once: They share an AST-walking infrastructure; landing them in one PR is cheaper than two.

Verify: Golden tests; round-trip fix application produces clean re-lint.

Task 12: `vox/require/justification-prose-required` + `vox/handler/panicking-builtin`

Files:

Create: crates/vox-code-audit/src/detectors/require/justification_prose.rs
Create: crates/vox-code-audit/src/detectors/handler/panicking_builtin.rs

Rules:

vox/require/justification-prose-required: A @require(<expr>) with > 1 operator in <expr> and no trailing prose comment ≥ 40 chars fires note. The fix appends a placeholder // because: … comment that the human fills in.

vox/handler/panicking-builtin: In any actor message handler body or workflow activity body, calls to known-panicking builtins (std.unwrap, std.expect, std.panic, std.unreachable, std.todo) fire warning. Suggested fix: replace with the Result-returning variant when one exists; otherwise wrap in match. Confidence varies.

Why these together: Both involve walking handler/activity bodies; shared infra.

Verify: Golden tests per rule. Anti-examples added to examples/golden/anti/.

Task 13: `@example` decorator + doctest/corpus unification

Files:

Modify: crates/vox-compiler/src/lower/decorators/example.rs (new decorator)
Modify: crates/vox-doc-pipeline/src/lib.rs — @example-tagged blocks become doctests
Modify: crates/vox-corpus/src/build.rs (or equivalent) — @example-tagged blocks become Mens corpus entries with mens_role: exemplar
Create: examples/golden/decorators/example.vox

Why: Today a “good example” lives in three places: a docstring, a separate examples/ file, and a Mens corpus entry. They drift. @example makes the source of truth singular: declare it once, the doctest runner verifies it compiles, the corpus pipeline includes it as training data.

Shape:

// vox:skip — @example decorator and named-arg syntax are planned features
@example("Resolve a secret and use it in an API call")
fn use_openai_key() -> Result[str, NetError] {
    let key = vox_secrets.resolve(SecretId::OpenAI)?;
    populi.complete(model: "gpt-4", api_key: key, prompt: "hi")
}

Verify: A vox-doc-pipeline --check run on a file with an @example decorator emits a doctest entry; a vox-corpus build --include-examples run includes the same fn in the corpus output. The decorator must error if applied to non-public fns (examples must be linkable from docs).

Task 14: `vox check --rationale-required` + suppression hygiene

Files:

Modify: crates/vox-cli/src/check.rs — add --rationale-required flag
Modify: crates/vox-code-audit/src/suppression.rs — parse // toestub-ignore(<id>) — <reason> and // vox:skip — <reason> (note the em-dash), fail the check if --rationale-required is set and any suppression lacks a reason of ≥ 20 chars
Modify: contracts/toestub/suppressions.v1.json schema — require owner (existing) and reason (existing) and add expires_at: (new, optional)

Why: Audit [A.28] applied to suppressions themselves. Without this, a suppression comment is itself a way to silently bypass a rule. With --rationale-required in CI, every suppression carries an auditable reason. The expires_at field lets transitional suppressions auto-expire.

Verify: Synthetic file with // vox:skip (no reason) → CI fails when --rationale-required is set; passes otherwise. Existing suppressions without reasons get a one-week grace period via a temporary allowlist.

Task 15: Negative-example corpus auto-built from compiler test fixtures

Files:

Create: crates/xtask/src/bin/gen_anti_corpus.rs
Modify: parser/typecheck test fixtures to wear a frontmatter declaration when their failure is a corpus-worthy anti-example
Create: examples/golden/anti/from-compiler-tests/ (output dir)

Why: Audit [A.24]. Today the compiler test fixtures and the Mens anti-corpus are separate. This task makes every parser test that asserts a diagnostic become a corpus anti-example automatically, with error_kind: and expected_diagnostic: derived from the test assertion.

Verify: Generator produces N anti-examples where N = count of compiler tests with the corpus frontmatter. CI fails if a compiler test asserts a diagnostic that doesn’t exist in the catalog.

Task 16: AGENTS.md backlinks + where-things-live update

Files:

Modify: AGENTS.md:
- §73–93 (Secret Management) — add “Enforced by vox/secret/env-get-shape.”
- §95–98 (Cryptography Policy) — “Enforced by vox/crypto/banned-crate-import and deny.toml.”
- §131–164 (Grammar Unification) — “Decorator-position discipline enforced by vox/decorator/position-mismatch.”
Modify: docs/src/architecture/where-things-live.md — add row for crates/vox-code-audit-macros (if Phase 1 didn’t), and one row per new detector module.

Verify: cargo run -p vox-doc-pipeline -- --check passes. where-things-live.md change is in same PR as detector adds (per Phase 1 Task 14 pattern).

Burn-down PRs (corpus hygiene)

After each detector lands as note or warning, a burn-down PR fixes existing violations before the next severity escalation. These PRs are not numbered tasks but are gates for severity escalation. One per detector that escalates.

Burn-down sequence (do in this order to minimize churn):

vox/llm/direct-provider-call — 0 expected violations (provider calls already route through populi); confirm and lock.
vox/secret/env-get-shape — audit existing env.get callsites; migrate any to vox_secrets.resolve.
vox/style/duplicate-prefix-name — across examples/golden/; autofixable.
vox/control-flow/question-mark-opportunity — across all .vox files; autofixable.
vox/doc/missing-adr-citation — for vox-actor-runtime/vox-orchestrator/vox-compiler pub fn only.

Risks specific to this phase

Risk	Mitigation
New detectors fire on existing corpus, blocking CI	Land each as `note` first; require burn-down PR before escalating to `warning` or `error`.
`Confidence::Certain` autofixes are wrong, an LLM applies them, code breaks	`Certain` is reserved for fixes that pass round-trip lint AND a parse-only verification. Add `cargo test -p vox-code-audit -- round_trip_certain_fixes` that applies every `Certain` fix in the test corpus and re-runs the detector.
`--for-llm` JSON shape becomes a public API and we want to change it	Schema field `"schema": "vox.lint.llm-report.v1"` allows a v2 alongside; v1 is kept until v3 lands.
Suppression hygiene (Task 14) breaks existing CI	Grace period: 30-day allowlist of pre-existing reasonless suppressions; CI warning until expiry, then error.
`@example` decorator (Task 13) overlaps with existing `examples/golden/` files	Migration: convert representative `examples/golden/` entries to `@example`-decorated fns living in test files; keep the rest as standalone golden examples (different role).

Phase 2 acceptance gate

Retrospective

Appended within 5 working days of phase completion.

Vox Language Rules — Phase 2: Lint Extension with Stable Diagnostic IDs (2026-05-09)

Phase 2 — vox-code-audit Lint Extension

Verification setup

Task 1: Detector trait + catalog populated with all 14 IDs

Task 2: vox/llm/direct-provider-call detector

Task 3: vox/secret/env-get-shape detector

Task 4: vox/crypto/banned-crate-import detector

Task 5: vox check --for-llm JSON mode

Task 6: vox check --explain <id> subcommand

Task 7: vox/doc/missing-adr-citation detector

Task 8: vox/decorator/position-mismatch detector

Task 9: vox/style/duplicate-prefix-name detector

Task 10: vox/style/long-range-coupling detector

Task 11: vox/control-flow/question-mark-opportunity + option-combinator-opportunity

Task 12: vox/require/justification-prose-required + vox/handler/panicking-builtin

Task 13: @example decorator + doctest/corpus unification

Task 14: vox check --rationale-required + suppression hygiene

Task 15: Negative-example corpus auto-built from compiler test fixtures

Task 16: AGENTS.md backlinks + where-things-live update

Burn-down PRs (corpus hygiene)

Risks specific to this phase

Phase 2 acceptance gate

Retrospective

Phase 2 — `vox-code-audit` Lint Extension

Task 2: `vox/llm/direct-provider-call` detector

Task 3: `vox/secret/env-get-shape` detector

Task 4: `vox/crypto/banned-crate-import` detector

Task 5: `vox check --for-llm` JSON mode

Task 6: `vox check --explain <id>` subcommand

Task 7: `vox/doc/missing-adr-citation` detector

Task 8: `vox/decorator/position-mismatch` detector

Task 9: `vox/style/duplicate-prefix-name` detector

Task 10: `vox/style/long-range-coupling` detector

Task 11: `vox/control-flow/question-mark-opportunity` + `option-combinator-opportunity`

Task 12: `vox/require/justification-prose-required` + `vox/handler/panicking-builtin`

Task 13: `@example` decorator + doctest/corpus unification

Task 14: `vox check --rationale-required` + suppression hygiene