Zeroize AuditSkill

---
name: zeroize-audit
description: "Detects missing zeroization of sensitive data in source code and identifies zeroization removed by compiler optimizations, with assembly-level analysis, and control-flow verification. Use for auditing C/C++/Rust code handling secrets, keys, passwords, or other sensitive data."
allowed-tools: Read Grep Glob Bash Write Task AskUserQuestion mcp__serena__activate_project mcp__serena__find_symbol mcp__serena__find_referencing_symbols mcp__serena__get_symbols_overview
---

# zeroize-audit — Claude Skill

## When to Use
- Auditing cryptographic implementations (keys, seeds, nonces, secrets)
- Reviewing authentication systems (passwords, tokens, session data)
- Analyzing code that handles PII or sensitive credentials
- Verifying secure cleanup in security-critical codebases
- Investigating memory safety of sensitive data handling

## When NOT to Use
- General code review without security focus
- Performance optimization (unless related to secure wiping)
- Refactoring tasks not related to sensitive data
- Code without identifiable secrets or sensitive values

---

## How to Run

On a request like "audit this crate for secrets left in memory" or "check that this C library actually wipes its keys":

1. **Collect inputs.** Map the request onto the Inputs table below (full schema: `{baseDir}/schemas/input.json`). `path` is required, plus at least one of `compile_db` (C/C++) or `cargo_manifest` (Rust); if neither is given or derivable from the repo, ask the user, because preflight stops the run without one. Leave all other fields at their defaults unless the user says otherwise.
2. **Read the orchestrator prompt, `{baseDir}/prompts/task.md`**, substituting the collected inputs for its `{{placeholder}}` values. You act as the orchestrator it describes: it defines state recovery, the phase loop, early termination, and error handling. Read `{baseDir}/prompts/system.md` alongside it for the shared working-directory layout and the agent error protocol every phase depends on.
3. **Execute its phase loop.** Run Phases 0-7 sequentially. Before each phase, read that phase's workflow file from `{baseDir}/workflows/phase-{N}-{name}.md` and follow its Preconditions, Instructions, State Update, and Error Handling sections. Each workflow specifies which agent to spawn via `Task` and with what parameters. Honor the per-phase skip conditions and the early-termination rules in task.md.
4. **Return the report** (Phase 8, inline): read `{workdir}/report/final-report.md` and return its contents as the skill output.

To resume an interrupted run: if a `workdir` is known from prior context, read `{workdir}/orchestrator-state.json` and continue from its `current_phase` instead of starting at Phase 0 (see the Recovery section of task.md).

---

## Purpose
Detect missing zeroization of sensitive data in source code and identify zeroization that is removed or weakened by compiler optimizations (e.g., dead-store elimination), with mandatory LLVM IR/asm evidence. Capabilities include:
- Assembly-level analysis for register spills and stack retention
- Data-flow tracking for secret copies
- Heap allocator security warnings
- Semantic IR analysis for loop unrolling and SSA form
- Control-flow graph analysis for path coverage verification
- Runtime validation test generation

## Scope
- Read-only against the target codebase (does not modify audited code; writes analysis artifacts to a temporary working directory).
- Produces a structured report (JSON).
- Requires valid build context (`compile_commands.json`) and compilable translation units.
- "Optimized away" findings only allowed with compiler evidence (IR/asm diff).

---

## Inputs

See `{baseDir}/schemas/input.json` for the full schema. Key fields:

| Field | Required | Default | Description |
|---|---|---|---|
| `path` | yes | — | Repo root |
| `compile_db` | no | `null` | Path to `compile_commands.json` for C/C++ analysis. Required if `cargo_manifest` is not set. |
| `cargo_manifest` | no | `null` | Path to `Cargo.toml` for Rust crate analysis. Required if `compile_db` is not set. |
| `config` | no | — | YAML defining heuristics and approved wipes |
| `opt_levels` | no | `["O0","O1","O2"]` | Optimization levels for IR comparison. O1 is the diagnostic level: if a wipe disappears at O1 it is simple DSE; O2 catches more aggressive eliminations. |
| `languages` | no | `["c","cpp","rust"]` | Languages to analyze |
| `max_tus` | no | `50` | Limit on translation units processed from compile DB |
| `mcp_mode` | no | `prefer` | `off`, `prefer`, or `require` — controls Serena MCP usage |
| `mcp_required_for_advanced` | no | `true` | Downgrade `SECRET_COPY`, `MISSING_ON_ERROR_PATH`, and `NOT_DOMINATING_EXITS` to `needs_review` when MCP is unavailable |
| `mcp_timeout_ms` | no | `10000` | Timeout budget for MCP semantic queries |
| `poc_categories` | no | all 11 exploitable | Finding categories for which to generate PoCs. C/C++ findings: all 11 categories supported. Rust findings: only `MISSING_SOURCE_ZEROIZE`, `SECRET_COPY`, and `PARTIAL_WIPE` are supported; other Rust categories are marked `poc_supported=false`. |
| `poc_output_dir` | no | `generated_pocs/` | Output directory for generated PoCs |
| `enable_asm` | no | `true` | Enable assembly emission and analysis (Step 8); produces `STACK_RETENTION`, `REGISTER_SPILL`. Auto-disabled if `emit_asm.sh` is missing. |
| `enable_semantic_ir` | no | `false` | Enable semantic LLVM IR analysis (Step 9); produces `LOOP_UNROLLED_INCOMPLETE` |
| `enable_cfg` | no | `false` | Enable control-flow graph analysis (Step 10); produces `MISSING_ON_ERROR_PATH`, `NOT_DOMINATING_EXITS` |
| `enable_runtime_tests` | no | `false` | Enable runtime test harness generation (Step 11) |

---

## Prerequisites

Before running, verify the following. Each has a defined failure mode.

**C/C++ prerequisites:**

| Prerequisite | Failure mode if missing |
|---|---|
| `compile_commands.json` at `compile_db` path | Fail fast — do not proceed |
| `clang` on PATH | Fail fast — IR/ASM analysis impossible |
| `uvx` on PATH (for Serena) | If `mcp_mode=require`: fail. If `mcp_mode=prefer`: continue without MCP; downgrade affected findings per Confidence Gating rules. |
| `{baseDir}/tools/extract_compile_flags.py` | Fail fast — cannot extract per-TU flags |
| `{baseDir}/tools/emit_ir.sh` | Fail fast — IR analysis impossible |
| `{baseDir}/tools/emit_asm.sh` | Warn and skip assembly findings (STACK_RETENTION, REGISTER_SPILL) |
| `{baseDir}/tools/mcp/check_mcp.sh` | Warn and treat as MCP unavailable |
| `{baseDir}/tools/mcp/normalize_mcp_evidence.py` | Warn and use raw MCP output |

**Rust prerequisites:**

| Prerequisite | Failure mode if missing |
|---|---|
| `Cargo.toml` at `cargo_manifest` path | Fail fast — do not proceed |
| `cargo check` passes | Fail fast — crate must be buildable |
| `cargo +nightly` on PATH | Fail fast — nightly required for MIR and LLVM IR emission |
| `uv` on PATH | Fail fast — required to run Python analysis scripts |
| `{baseDir}/tools/validate_rust_toolchain.sh` | Warn — run preflight manually. Checks all tools, scripts, nightly, and optionally `cargo check`. Use `--json` for machine-readable output, `--manifest` to also validate the crate builds. |
| `{baseDir}/tools/emit_rust_mir.sh` | Fail fast — MIR analysis impossible (`--opt`, `--crate`, `--bin/--lib` supported; `--out` can be file or directory) |
| `{baseDir}/tools/emit_rust_ir.sh` | Fail fast — LLVM IR analysis impossible (`--opt` required; `--crate`, `--bin/--lib` supported; `--out` must be `.ll`) |
| `{baseDir}/tools/emit_rust_asm.sh` | Warn and skip assembly findings (`STACK_RETENTION`, `REGISTER_SPILL`). Supports `--opt`, `--crate`, `--bin/--lib`, `--target`, `--intel-syntax`; `--out` can be `.s` file or directory. |
| `{baseDir}/tools/diff_rust_mir.sh` | Warn and skip MIR-level optimization comparison. Accepts 2+ MIR files, normalizes, diffs pairwise, and reports first opt level where zeroize/drop-glue patterns disappear. |
| `{baseDir}/tools/scripts/semantic_audit.py` | Warn and skip semantic source analysis |
| `{baseDir}/tools/scripts/find_dangerous_apis.py` | Warn and skip dangerous API scan |
| `{baseDir}/tools/scripts/check_mir_patterns.py` | Warn and skip MIR analysis |
| `{baseDir}/tools/scripts/check_llvm_patterns.py` | Warn and skip LLVM IR analysis |
| `{baseDir}/tools/scripts/check_rust_asm.py` | Warn and skip Rust assembly analysis (`STACK_RETENTION`, `REGISTER_SPILL`, drop-glue checks). Dispatches to `check_rust_asm_x86.py` (production) or `check_rust_asm_aarch64.py` (**EXPERIMENTAL** — AArch64 findings require manual verification). |
| `{baseDir}/tools/scripts/check_rust_asm_x86.py` | Required by `check_rust_asm.py` for x86-64 analysis; warn and skip if missing |
| `{baseDir}/tools/scripts/check_rust_asm_aarch64.py` | Required by `check_rust_asm.py` for AArch64 analysis (**EXPERIMENTAL**); warn and skip if missing |

**Common prerequisite:**

| Prerequisite | Failure mode if missing |
|---|---|
| `{baseDir}/tools/generate_poc.py` | Fail fast — PoC generation is mandatory |

---

## Approved Wipe APIs

The following are recognized as valid zeroization. Configure additional entries in `{baseDir}/configs/`.

**C/C++**
- `explicit_bzero`
- `memset_s`
- `SecureZeroMemory`
- `OPENSSL_cleanse`
- `sodium_memzero`
- Volatile wipe loops (pattern-based; see `volatile_wipe_patterns` in `{baseDir}/configs/default.yaml`)
- In IR: `llvm.memset` with volatile flag, volatile stores, or non-elidable wipe call

**Rust**
- `zeroize::Zeroize` trait (`zeroize()` method)
- `Zeroizing<T>` wrapper (drop-based)
- `ZeroizeOnDrop` derive macro

---

## Finding Capabilities

Findings are grouped by required evidence. Only attempt findings for which the required tooling is available.

| Finding ID | Description | Requires | PoC Support |
|---|---|---|---|
| `MISSING_SOURCE_ZEROIZE` | No zeroization found in source | Source only | Yes (C/C++ + Rust) |
| `PARTIAL_WIPE` | Incorrect size or incomplete wipe | Source only | Yes (C/C++ + Rust) |
| `NOT_ON_ALL_PATHS` | Zeroization missing on some control-flow paths (heuristic) | Source only | Yes (C/C++ only) |
| `SECRET_COPY` | Sensitive data copied without zeroization tracking | Source + MCP preferred | Yes (C/C++ + Rust) |
| `INSECURE_HEAP_ALLOC` | Secret uses insecure allocator (malloc vs. secure_malloc) | Source only | Yes (C/C++ only) |
| `OPTIMIZED_AWAY_ZEROIZE` | Compiler removed zeroization | IR diff required (never source-only) | Yes |
| `STACK_RETENTION` | Stack frame may retain secrets after return | Assembly required (C/C++); LLVM IR `alloca`+`lifetime.end` evidence (Rust); assembly corroboration upgrades to `confirmed` | Yes (C/C++ only) |
| `REGISTER_SPILL` | Secrets spilled from registers to stack | Assembly required (C/C++); LLVM IR `load`+call-site evidence (Rust); assembly corroboration upgrades to `confirmed` | Yes (C/C++ only) |
| `MISSING_ON_ERROR_PATH` | Error-handling paths lack cleanup | CFG or MCP required | Yes |
| `NOT_DOMINATING_EXITS` | Wipe doesn't dominate all exits | CFG or MCP required | Yes |
| `LOOP_UNROLLED_INCOMPLETE` | Unrolled loop wipe is incomplete | Semantic IR required | Yes |

---

## Agent Architecture

The analysis pipeline uses 11 agents across 8 phases, invoked by the orchestrator (`{baseDir}/prompts/task.md`) via `Task`. Agents write persistent finding files to a shared working directory (`/tmp/zeroize-audit-{run_id}/`), enabling parallel execution and protecting against context pressure.

| Agent | Phase | Purpose | Output Directory |
|---|---|---|---|
| `0-preflight` | Phase 0 | Preflight checks (tools, toolchain, compile DB, crate build), config merge, workdir creation, TU enumeration | `{workdir}/` |
| `1-mcp-resolver` | Phase 1, Wave 1 (C/C++ only) | Resolve symbols, types, and cross-file references via Serena MCP | `mcp-evidence/` |
| `2-source-analyzer` | Phase 1, Wave 2a (C/C++ only) | Identify sensitive objects, detect wipes, validate correctness, data-flow/heap | `source-analysis/` |
| `2b-rust-source-analyzer` | Phase 1, Wave 2b (Rust only, parallel with 2a) | Rustdoc JSON trait-aware analysis + dangerous API grep | `source-analysis/` |
| `3-tu-compiler-analyzer` | Phase 2, Wave 3 (C/C++ only, N parallel) | Per-TU IR diff, assembly, semantic IR, CFG analysis | `compiler-analysis/{tu_hash}/` |
| `3b-rust-compiler-analyzer` | Phase 2, Wave 3R (Rust only, single agent) | Crate-level MIR, LLVM IR, and assembly analysis | `rust-compiler-analysis/` |
| `4-report-assembler` | Phase 3 (interim) + Phase 6 (final) | Collect findings from all agents, apply confidence gates; merge PoC results and produce final report | `report/` |
| `5-poc-generator` | Phase 4 | Craft bespoke proof-of-concept programs (C/C++: all categories; Rust: MISSING_SOURCE_ZEROIZE, SECRET_COPY, PARTIAL_WIPE) | `poc/` |
| `5b-poc-validator` | Phase 5 | Compile and run all PoCs | `poc/` |
| `5c-poc-verifier` | Phase 5 | Verify each PoC proves its claimed finding | `poc/` |
| `6-test-generator` | Phase 7 (optional) | Generate runtime validation test harnesses | `tests/` |

The orchestrator reads one per-phase workflow file from `{baseDir}/workflows/` at a time, and maintains `orchestrator-state.json` for recovery after context compression. Agents receive configuration by file path (`config_path`), not by value.

### Execution flow

```
Phase 0: 0-preflight agent — Preflight + config + create workdir + enumerate TUs
           → writes orchestrator-state.json, merged-config.yaml, preflight.json
Phase 1: Wave 1:  1-mcp-resolver              (skip if mcp_mode=off OR language_mode=rust)
         Wave 2a: 2-source-analyzer           (C/C++ only; skip if no compile_db)  ─┐ parallel
         Wave 2b: 2b-rust-source-analyzer     (Rust only; skip if no cargo_manifest) ─┘
Phase 2: Wave 3:  3-tu-compiler-analyzer x N  (C/C++ only; parallel per TU)
         Wave 3R: 3b-rust-compiler-analyzer   (Rust only; single crate-level agent)
Phase 3: Wave 4:  4-report-assembler          (mode=interim → findings.json; reads all agent outputs)
Phase 4: Wave 5:  5-poc-generator             (C/C++: all categories; Rust: MISSING_SOURCE_ZEROIZE, SECRET_COPY, PARTIAL_WIPE; other Rust findings: poc_supported=false)
Phase 5: PoC Validation & Verification
           Step 1: 5b-poc-validator agent      (compile and run all PoCs)
           Step 2: 5c-poc-verifier agent       (verify each PoC proves its claimed finding)
           Step 3: Orchestrator presents verification failures to user via AskUserQuestion
           Step 4: Orchestrator merges all results into poc_final_results.json
Phase 6: Wave 6: 4-report-assembler           (mode=final → merge PoC results, final-report.md)
Phase 7: Wave 7: 6-test-generator             (optional)
Phase 8: Orchestrator — Return final-report.md
```

## Cross-Reference Convention

IDs are namespaced per agent to prevent collisions during parallel execution:

| Entity | Pattern | Assigned By |
|---|---|---|
| Sensitive object (C/C++) | `SO-0001`–`SO-4999` | `2-source-analyzer` |
| Sensitive object (Rust) | `SO-5000`–`SO-9999` (Rust namespace) | `2b-rust-source-analyzer` |
| Source finding (C/C++) | `F-SRC-NNNN` | `2-source-analyzer` |
| Source finding (Rust) | `F-RUST-SRC-NNNN` | `2b-rust-source-analyzer` |
| IR finding (C/C++) | `F-IR-{tu_hash}-NNNN` | `3-tu-compiler-analyzer` |
| ASM finding (C/C++) | `F-ASM-{tu_hash}-NNNN` | `3-tu-compiler-analyzer` |
| CFG finding | `F-CFG-{tu_hash}-NNNN` | `3-tu-compiler-analyzer` |
| Semantic IR finding | `F-SIR-{tu_hash}-NNNN` | `3-tu-compiler-analyzer` |
| Rust MIR finding | `F-RUST-MIR-NNNN` | `3b-rust-compiler-analyzer` |
| Rust LLVM IR finding | `F-RUST-IR-NNNN` | `3b-rust-compiler-analyzer` |
| Rust assembly finding | `F-RUST-ASM-NNNN` | `3b-rust-compiler-analyzer` |
| Translation unit | `TU-{hash}` | Orchestrator |
| Final finding | `ZA-NNNN` | `4-report-assembler` |

Every finding JSON object includes `related_objects`, `related_findings`, and `evidence_files` fields for cross-referencing between agents.

---

## Detection Strategy

Analysis runs in two phases. For complete step-by-step guidance, see `{baseDir}/references/detection-strategy.md`.

| Phase | Steps | Findings produced | Required tooling |
|---|---|---|---|
| Phase 1 (Source) | 1–6 | `MISSING_SOURCE_ZEROIZE`, `PARTIAL_WIPE`, `NOT_ON_ALL_PATHS`, `SECRET_COPY`, `INSECURE_HEAP_ALLOC` | Source + compile DB |
| Phase 2 (Compiler) | 7–12 | `OPTIMIZED_AWAY_ZEROIZE`, `STACK_RETENTION`*, `REGISTER_SPILL`*, `LOOP_UNROLLED_INCOMPLETE`†, `MISSING_ON_ERROR_PATH`‡, `NOT_DOMINATING_EXITS`‡ | `clang`, IR/ASM tools |

\* requires `enable_asm=true` (default)
† requires `enable_semantic_ir=true`
‡ requires `enable_cfg=true`

For Rust, `{baseDir}/references/rust-zeroization-patterns.md` catalogues 40 named anti-patterns, keyed to the script that detects each one: Section A for rustdoc-JSON semantics (`semantic_audit.py`), Section B for dangerous APIs (`find_dangerous_apis.py`), and Section C for MIR/LLVM IR/assembly (`check_mir_patterns.py`, `check_llvm_patterns.py`, `check_rust_asm.py`). Read the relevant section when triaging a Rust finding, writing its fix recommendation, or deciding whether a hand-spotted pattern is already covered.

Two limits on how far that reference goes. The 34 entries in Sections A-C are what the scripts detect today; Section D's six are known gaps no script covers, so treat those as unaudited rather than clean. Sections A and C are also partial — the scripts emit some classes with no entry — so a finding that matches no catalogued pattern is still a finding, carrying whatever evidence the script produced.

---


## Output Format

Each run produces two outputs:

1. **`final-report.md`** — Comprehensive markdown report (primary human-readable output)
2. **`findings.json`** — Structured JSON matching `{baseDir}/schemas/output.json` (for machine consumption and downstream tools)

### Markdown Report Structure

The markdown report (`final-report.md`) contains these sections:

- **Header**: Run metadata (run_id, timestamp, repo, compile_db, config summary)
- **Executive Summary**: Finding counts by severity, confidence, and category
- **Sensitive Objects Inventory**: Table of all identified objects with IDs, types, locations
- **Findings**: Grouped by severity then confidence. Each finding includes location, object, all evidence (source/IR/ASM/CFG), compiler evidence details, and recommended fix
- **Superseded Findings**: Source findings replaced by CFG-backed findings
- **Confidence Gate Summary**: Downgrades applied and overrides rejected
- **Analysis Coverage**: TUs analyzed, agent success/failure, features enabled, and any Section D patterns the crate uses that no script audits
- **Appendix: Evidence Files**: Mapping of finding IDs to evidence file paths

### Structured JSON

The `findings.json` file follows the schema in `{baseDir}/schemas/output.json`. Each `Finding` object:

```json
{
  "id": "ZA-0001",
  "category": "OPTIMIZED_AWAY_ZEROIZE",
  "severity": "high",
  "confidence": "confirmed",
  "language": "c",
  "file": "src/crypto.c",
  "line": 42,
  "symbol": "key_buf",
  "evidence": "store volatile i8 0 count: O0=32, O2=0 — wipe eliminated by DSE",
  "compiler_evidence": {
    "opt_levels": ["O0", "O2"],
    "o0": "32 volatile stores targeting key_buf",
    "o2": "0 volatile stores (all eliminated)",
    "diff_summary": "All volatile wipe stores removed at O2 — classic DSE pattern"
  },
  "suggested_fix": "Replace memset with explicit_bzero or add compiler_fence(SeqCst) after the wipe",
  "poc": {
    "file": "generated_pocs/ZA-0001.c",
    "makefile_target": "ZA-0001",
    "compile_opt": "-O2",
    "requires_manual_adjustment": false,
    "validated": true,
    "validation_result": "exploitable"
  }
}
```

See `{baseDir}/schemas/output.json` for the full schema and enum values.

---

## Confidence Gating

### Evidence thresholds

A finding requires at least **2 independent signals** to be marked `confirmed`. With 1 signal, mark `likely`. With 0 strong signals (name-pattern match only), mark `needs_review`.

Signals include: name pattern match, type hint match, explicit annotation, IR evidence, ASM evidence, MCP cross-reference, CFG evidence, PoC validation.

### PoC validation as evidence signal

Every finding is validated against a bespoke PoC. After compilation and execution, each PoC is also verified to ensure it actually tests the claimed vulnerability. The combined result is an evidence signal:

| PoC Result | Verified | Impact |
|---|---|---|
| Exit 0 (exploitable) | Yes | Strong signal — can upgrade `likely` to `confirmed` |
| Exit 1 (not exploitable) | Yes | Downgrade severity to `low` (informational); retain in report |
| Exit 0 or 1 | No (user accepted) | Weaker signal — note verification failure in evidence |
| Exit 0 or 1 | No (user rejected) | No confidence change; annotate as `rejected` |
| Compile failure / no PoC | — | No confidence change; annotate in evidence |

### MCP unavailability downgrade

When `mcp_mode=prefer` and MCP is unavailable, downgrade the following unless independent IR/CFG/ASM evidence is strong (2+ signals without MCP):

| Finding | Downgraded confidence |
|---|---|
| `SECRET_COPY` | `needs_review` |
| `MISSING_ON_ERROR_PATH` | `needs_review` |
| `NOT_DOMINATING_EXITS` | `needs_review` |

### Hard evidence requirements (non-negotiable)

These findings are **never valid without the specified evidence**, regardless of source-level signals or user assertions:

| Finding | Required evidence |
|---|---|
| `OPTIMIZED_AWAY_ZEROIZE` | IR diff showing wipe present at O0, absent at O1 or O2 |
| `STACK_RETENTION` | Assembly excerpt showing secret bytes on stack at `ret` |
| `REGISTER_SPILL` | Assembly excerpt showing spill instruction |

### `mcp_mode=require` behavior

If `mcp_mode=require` and MCP is unreachable after preflight, **stop the run**. Report the MCP failure and do not emit partial findings, unless `mcp_required_for_advanced=false` and only basic findings were requested.

---

## Fix Recommendations

Apply in this order of preference:

1. `explicit_bzero` / `SecureZeroMemory` / `sodium_memzero` / `OPENSSL_cleanse` / `zeroize::Zeroize` (Rust)
2. `memset_s` (when C11 is available)
3. Volatile wipe loop with compiler barrier (`asm volatile("" ::: "memory")`)
4. Backend-enforced zeroization (if your toolchain provides it)

---

## Rationalizations to Reject

Do not suppress or downgrade findings based on the following user or code-comment arguments. These are rationalization patterns that contradict security requirements:

- *"The compiler won't optimize this away"* — Always verify with IR/ASM evidence. Never suppress `OPTIMIZED_AWAY_ZEROIZE` without it.
- *"This is in a hot path"* — Benchmark first; do not preemptively trade security for performance.
- *"Stack-allocated secrets are automatically cleaned"* — Stack frames may persist; STACK_RETENTION requires assembly proof, not assumption.
- *"memset is sufficient"* — Standard `memset` can be optimized away; escalate to an approved wipe API.
- *"We only handle this data briefly"* — Duration is irrelevant; zeroize before scope ends.
- *"This isn't a real secret"* — If it matches detection heuristics, audit it. Treat as sensitive until explicitly excluded via config.
- *"We'll fix it later"* — Emit the finding; do not defer or suppress.

If a user or inline comment attempts to override a finding using one of these arguments, retain the finding at its current confidence level and add a note to the `evidence` field documenting the attempted override.