Tech-Debt Audit

Audit ByroRedux for accumulated technical debt: code that compiles, passes tests and ships, but raises the
cost of every future change. Correctness bugs belong to other audits; this one hunts decay since the last
cleanup pass.

Every dimension below is a DISCOVERY RECIPE, not a finding list. Instances churn between audits, so each
dimension hands you a command that enumerates current instances plus a triage rule. There is no hardcoded
instance list here on purpose — re-run the recipe and report what it surfaces today. Where a cargo test /
CI gate already enforces an invariant, the dimension names the guard and aims at what it cannot see.

Architecture: Orchestrator. Each dimension runs as a Task agent (max 3 concurrent).

See .claude/commands/_audit-common.md for layout, crate roster, methodology, dedup, severity and finding
format. Young code has had the fewest sweeps — find it with
git log --diff-filter=A --since=<last-report-date> --name-only --format= -- 'crates/*/Cargo.toml' 'tools/*/Cargo.toml'
plus the crates the last report never names (for c in crates/*/; do grep -q "$c" <last-report> || echo "$c"; done;
the 2026-10-08 report still named none of sdk, mod-runtime, menuxml, save, hkx, spt, pex,
sfmaterial, audio, debug-ui, game-detect — 21 crates in all — nor three of the four tools/). File real findings there; do not
just note the crate is young. crates/cxx-bridge and crates/platform are deliberate placeholders owned here:
check they have not grown a second job or silent consumers, not that they are small.

Parameters (from $ARGUMENTS)

• --focus <dimensions>: comma-separated dimension numbers (e.g. 1,3,5). Default: all 9.
• --depth shallow|deep: shallow = surface counts + worst offenders; deep = per-instance triage with a concrete fix proposal. Default deep.

Extra Per-Finding Fields

• Dimension: one of the 9 below.
• Age (when relevant): commit hash + date the debt landed (git log -L / git blame).
• Effort: trivial (≤30 min) | small (≤2 h) | medium (≤1 day) | large (>1 day, decompose first).
• ID convention: TD<dim>-<date>-NN (e.g. TD3-2026-09-05-02 = Dim 3 Stale Documentation, finding 2 of that report).

Severity for Tech Debt

Tech-debt findings default to LOW (see _audit-severity.md). Promote only on amplification:

Promotion Trigger	Floor
Duplicated logic with divergent bug-fix history (one branch fixed, the other regressed)	MEDIUM
unimplemented!() / todo!() / panic!("not …") reachable from a shipped CLI flag or smoke test	MEDIUM
#[ignore]d test that guards a fix from a closed CRITICAL/HIGH issue	MEDIUM
Stale doc/audit baseline that misled an audit in the last 90 days	MEDIUM
Magic number that would silently over/underflow under documented use	HIGH
Stale GpuCamera/GpuInstance/GpuMaterial size in a doc comment (lockstep-drift bait)	MEDIUM

Phase 1: Setup

1. Parse $ARGUMENTS. 2. mkdir -p /tmp/audit/tech-debt.
2. Dedup baseline:
   gh issue list --repo matiaszanolli/ByroRedux --limit 500 --state all --label tech-debt --json number,title,state > /tmp/audit/tech-debt/issues_all.json
3. Scan docs/audits/ for prior AUDIT_TECH_DEBT_*.md (diff direction, not re-litigation).
4. Production-LOC helper (Dim 1's subject is production complexity, not file length — a file long from bulk
   inline tests is not the debt this dimension hunts). Source the checked-in script and run its self-test first;
   a failure means no figure below it can be trusted:

   source .claude/commands/audit-tech-debt/prod_loc.sh
   prod_loc_self_test        # must print "prod_loc self-test: ok"

   prod_loc <file> reports 0 for pure-test files (tests.rs, *_tests.rs, anything under tests/ — their
   #[cfg(test)] mod gate lives in the parent) and otherwise total LOC minus every #[cfg(test)]-gated braced
   item, counting braces on code only (comments, strings, raw strings and char literals stripped, #4336).
5. Snapshot totals so the next audit can diff. tools/ is in scope (four workspace binaries); tools/nifskope is
   vendored and pruned:

   {
     echo "markers (TODO/FIXME/HACK/XXX/TBD/WIP/KLUDGE): $(grep -RInE '(TODO|FIXME|HACK|XXX|TBD|WIP|KLUDGE)\b' crates byroredux tools --exclude-dir=nifskope | wc -l)"
     echo "allow(dead_code):      $(grep -RInE 'allow\(dead_code\)' crates byroredux tools --exclude-dir=nifskope | wc -l)"
     echo "unimplemented!/todo!(): $(grep -RInE 'unimplemented!|todo!\(\)' crates byroredux tools --exclude-dir=nifskope | wc -l)"
     echo "#[ignore] tests:        $(grep -RInE '^[[:space:]]*#\[ignore' --include='*.rs' crates byroredux tools --exclude-dir=nifskope | wc -l)"
     echo "files >2000 production LOC: $(for f in $(find crates byroredux tools -name '*.rs' -not -path 'tools/nifskope/*'); do echo "$(prod_loc "$f")"; done | awk '$1>2000' | wc -l)"
     echo "test files >2000 total LOC (lower priority, separate bucket): $(find crates byroredux tools -name '*.rs' -not -path 'tools/nifskope/*' -exec wc -l {} + | awk '$1>2000 && $2!="total"' | wc -l)"
   } > /tmp/audit/tech-debt/baseline.txt

   Measured 2026-10-12 at 0d698d16c (diff direction only, re-run, never quote): markers 24 (15 are XXXX false positives, 2 the
   FO4 HACK FourCC — Dim 5), allow(dead_code) 44 (45 on 2026-10-10; the 11 added since 2026-10-08 are cfg(test)
   TypeId-only markers in crates/core/src/ecs/lock_tracker.rs, c5113f0d9 — not debt), unimplemented!/todo!() 0 (a
   fresh hit is notable), #[ignore] 299 (287 on 2026-10-10; tools-inclusive — earlier reports scoped to crates+byroredux
   read lower),
   production >2000 LOC: 4 (shader_constants_data.rs is a data table, Dim 1; byroredux/src/components.rs;
   crates/scripting/src/fragment/effects.rs; byroredux/src/save_io.rs, new — it crossed to 2011. byroredux/src/env_translate/
   left the list: it was one file at 1881 → 2186 across 2026-10-05..11 on the #5374/#5388/#5423/#5424 climate/LOD fixes and #5584 split it
   per concern into lod / climate / water / sky_weather / fallback, none over 700 production LOC; watch-list within 5%:
   crates/nif/src/import/types.rs (1997), crates/audio/src/lib.rs, crates/plugin/src/esm/cell/mod.rs,
   crates/scripting/src/translate/effects.rs, crates/scripting/examples/mq101_conformance.rs), test-heavy >2000: 77.
   A raw whole-repo grep for #[ignore] also matches markdown prose — keep --include='*.rs'.

Phase 2: Dimension Agents

Ordered by debt impact: complexity and duplication compound across every edit; doc/audit rot misdirects the next
audit; markers and dead code are cheap. Recent yield (grep -h '\*\*Dimension\*\*' docs/audits/AUDIT_TECH_DEBT_*.md):
Dims 4, 3, 1, 8 dominate (then 2, 9); 5 and 6 rarely surface real debt. Each agent writes /tmp/audit/tech-debt/dim_<N>.md.

Dimension 1: File / Function / Module Complexity

Paths: all .rs under crates/, byroredux/, tools/
First step: the two-bucket command below

An oversized file taxes every edit, review and merge. Threshold is 2000 production LOC. Two buckets:

# Primary — the dimension's actual subject. File real findings from this.
for f in $(find crates byroredux tools -name '*.rs' -not -path 'tools/nifskope/*'); do
    p=$(prod_loc "$f"); [ "$p" -gt 2000 ] && echo -e "$p\t$f"
done | sort -rn
# Secondary — test-heavy files, report but do not auto-file: escalate one only if its OWN prod_loc also crosses 2000.
find crates byroredux tools -name '*.rs' -not -path 'tools/nifskope/*' -exec wc -l {} + | awk '$1>2000 && $2!="total"' | sort -rn

The recipe output is authoritative. A file that once split can re-cross (a previously-split file that grew back is a
live finding); a file that dips under by incidental shrinkage is still a candidate — only an actual split removes one;
and files within ~5% of the line (check the snapshot) are worth naming as watch-list rows. Function-level splits inside
a file and file-level splits are independent signals — a register_*_systems / phase-helper extraction does not move
its host file across the line.

Per file, propose a split AXIS by responsibility, chosen after reading the file's internal structure — never from a
name or a tag. Precedent: a metadata enum tag (ExtenderFamily) touching 30 of 3759 lines was not the seam; the real
axis was the four-layer service stack (routes → declarations → source aliases → runtime adapters) with one service
holding over half the lines (#3851). Splits that worked: Vulkan context/ and volumetrics by
construct-vs-record-vs-teardown; boot.rs into one file per registration stage; runtime.rs into one file per
impl <wit>::Host block. Data tables (FourCC → id maps, shader_constants_data.rs) want a static table or
generation, not a split. Vulkan-recording splits are render-pass-adjacent — read
feedback_speculative_vulkan_fixes before proposing barrier/order changes.

Also flag: functions >200 LOC (propose extraction); match arms >50 cases (want a lookup table); nesting depth >5;
a mod.rs / lib.rs with >20 pub use (two jobs). cargo +nightly clippy --all-targets -- -W clippy::cognitive_complexity
if available, else inspect the worst offenders.

Dimension 2: Logic Duplication

Paths: sibling-file families under crates/nif/src/blocks/, crates/plugin/src/esm/records/, crates/renderer/src/vulkan/, byroredux/src/cell_loader/, byroredux/src/systems/
First step: ls crates/nif/src/blocks/*.rs crates/plugin/src/esm/records/*.rs crates/plugin/src/esm/records/*/*.rs crates/renderer/src/vulkan/*.rs byroredux/src/cell_loader/*.rs, then read two siblings side by side

CLAUDE.md global policy: improve existing code, never duplicate logic. Every finding names a concrete
consolidation site (an existing helper to extend, or the new one and its callers).
Look for: block-parser scaffolding (header → field → fixup) repeated across crates/nif/src/blocks/; texture-upload
chains (BC1/BC3/BC5/RGBA) and image-layout barrier sequences repeated per pass in crates/renderer/src/vulkan/
(vulkan/image.rs GpuImage and the shared barrier helpers are the consolidation targets — a pass that still
hand-rolls create/bind/destroy is a candidate); vk::WriteDescriptorSet builder boilerplate; ESM sub-record parse loops
across crates/plugin/src/esm/records/; AI-procedure systems sharing scaffolding; near-identical per-game HUD/profile
tables; source-scan helpers (production_text, production_text_or_whole, strip_test_modules, count_identifier) have ONE
home, crates/core/src/source_scan.rs (#5100, 655b317c9; #5469, 334458b3f routed the ~40 surviving inline cuts through it) — a
re-copied cut in a test module is a regression; no_source_scan_cuts_with_the_weak_cfg_test_needle
(byroredux/src/workspace_hygiene_tests.rs) fails a new weak #[cfg(test)] find/split outside its allowlist. The Z-up→Y-up conversion has one home, crates/core/src/math/coord.rs (zup_to_yup_pos,
zup_to_yup_quat_wxyz); re-exports elsewhere are not leaks — a genuine reimplementation would live outside that file.

Dimension 3: Stale Documentation & Comments

Paths: docs/, ROADMAP.md, HISTORY.md, README.md, CLAUDE.md, doc comments across the tree
First step: .claude/commands/_audit-validate.sh

Doc rot misleads the next reader and the next audit. Run the gate first: _audit-validate.sh resolves every
backticked path in the audit skills AND docs/engine/*.md against the tree (fatal on STALE) and lists backticked
symbols found in no tracked .rs file (advisory — clear it, do not learn to scroll past it). STALE refs are
auto-eligible findings (trivial). Guards that already police a class — aim at what they cannot see:

• GPU-struct sizes: gpu_material_size_claims (crates/renderer/src/vulkan/material_tests.rs) scans crates,
  byroredux, tools, docs/engine, .claude/commands and the top-level status docs for a stale byte count of every
  size-pinned struct — GpuMaterial, GpuLight, GpuTerrainTile, GpuInstance, GpuCamera since #5203
  (pinned_sizes(); plus bindings_glsl_states_the_real_struct_size). It CANNOT see Vertex::SIZE (a size written
  without the type name) or GpuWaterParams-class structs outside pinned_sizes() — cross-check those against the
  layout tests, never the prose:
  grep -rn "fn gpu_.*_is_[0-9]\+_bytes\|size_of::<Gpu" crates/renderer/src/vulkan/. A GPU-size claim found stale in
  prose is MEDIUM (severity table).
• Deleted Material::classify_pbr: no_source_file_frames_the_deleted_classify_pbr_as_live
  (byroredux/src/workspace_hygiene_tests.rs) fails a doc that names it as live. The survivors are the free fn
  classify_pbr_keyword and Material::resolve_pbr.
• Beyond the guards, sweep: doc comments naming renamed/deleted symbols (git log --diff-filter=D --since=<last> then
  grep each deleted name); ROADMAP.md milestones "in progress" whose issues are closed (or vice versa) — cross-check
  git log / gh issue; docs/feature-matrix.md rows against the crate that implements them (git log --grep M45,
  …) — it is a status floor, so flag any row contradicting shipped code; HISTORY.md entries referencing reverted work;
  README.md / CLAUDE.md command examples whose flags or paths changed; per-game compat numbers in docs/engine/game-compatibility.md
  vs ROADMAP.md's matrix (ROADMAP is the single home); docs/legacy/ Gamebryo source paths that moved;
  crates/renderer/shaders/triangle.frag comments quoting outdated struct sizes.
• docs/engine/*.md are the "authoritative" references audits are told to believe (_audit-common.md): beyond the gate's
  path/symbol check, sample their numbers and lists against code — stage order and lock order (ecs.md), VRAM/pool figures
  (memory-budget.md), shader binding tables (shader-pipeline.md), dependency and crate lists — a doc that misled an audit
  in the last 90 days is MEDIUM.
• Path convention: a backticked path or snake_case symbol in an audit skill asserts "exists right now"; deleted or
  forward-looking names go in italics or plain text (_audit-common.md Path-Reference Convention).

Dimension 4: Audit-Finding Rot

Paths: .claude/commands/, docs/audits/
First step: .claude/commands/_audit-validate.sh then ls .claude/commands/ docs/audits/ | tail -40

Stale baselines actively misdirect future audits.

• STALE refs the gate prints in other audit skills → Dim 4 findings (trivial); advisory symbols in skills likewise.
• Numeric claims in skills and _audit-common.md (LOC figures, struct sizes, counts, "all N dimensions") — the gate
  cannot see numbers: re-measure a sample (wc -l, prod_loc, grep -c) and flag any off by >20%. Layout-row LOC
  drift has been the single most repeated finding here.
• "Existing: #NNN" / "open issue" callouts in skills where the issue is now CLOSED: gh issue view N --json state.
• Dimension cross-references between skills ("/audit-x Dim N") that no longer name that dimension.
• docs/audits/ reports older than 90 days whose CRITICAL/HIGH findings have no GitHub trace. Caveat: reports
  dated before 2026-06-07 predate /audit-publish, so a missing issue is not evidence a finding is open (#3875,
  _audit-common.md Deduplication) — verify against code and say which you checked.
• Do NOT flag .claude/issues/<N>/ISSUE.md "Status: Open" drift: local issue files are immutable snapshots (TD10-001 / #1156);
  gh issue view <N> --json state is authoritative.

Dimension 5: Stale Markers (TODO / FIXME / HACK / XXX / TBD)

Paths: crates/, byroredux/, tools/ (not tools/nifskope), crates/renderer/shaders/
First step: the two greps below

grep -RInE '(TODO|FIXME|HACK|XXX|TBD|WIP|KLUDGE)\b' crates byroredux tools --exclude-dir=nifskope
grep -RInE '(TODO|FIXME|HACK|XXX|TBD|WIP|KLUDGE)\b' crates/renderer/shaders/

Keep the token list identical in both commands (an asymmetry once hid every shader FIXME, #3877); TBD is in it because
the codebase uses it. tools/ (byro-dbg, byro-launcher, byro-detect, texture-upscale) is first-party and in scope.
Triage each (skip markers <30 days old unless they name a closed issue): git blame for age — over 6 months gets
reported; does it name an issue, and is that issue still open (closed issue + live marker = "marker outlived its driver");
does it name a milestone now complete per ROADMAP.md; a // TODO: implement on a path reachable from a shipped CLI flag
promotes (severity table).
False positives: XXXX, the ESM extended-size sub-record tag (key on comment content — "references the ESM XXXX
escape" — not a file list; reader.rs, cell/wrld.rs, the b"XXXX" test sentinels in records/misc/magic.rs and the
FourCC table doc in sdk/src/compatibility/storage_util/mod.rs all qualify); HACK, the FO4 terminal-record FourCC
(crates/plugin/tests/parse_real_esm.rs); a // FIXME quoting a reference implementation's
own FIXME (crates/bgsm/src/bgem.rs, bs_geometry.rs) documents upstream; a TBD that records an unresolved format
semantic together with its own resolution (the FNV WEAP DNAM arm in crates/plugin/src/esm/records/items.rs) is a
documented unknown, not a stale marker.
Must-not-delete: the third-party attribution block atop crates/renderer/shaders/triangle.frag (GLSL-PathTracer MIT
notice + Burley citation, ~first 30 lines). Flag any edit that strips it — MIT requires the notice to travel with the code.

Dimension 6: Stub & Placeholder Implementations

Paths: crates/, byroredux/, tools/
First step: the two greps below

grep -RInE 'unimplemented!|todo!\(\)|panic!\("not ' crates byroredux tools --exclude-dir=nifskope
grep -RInE '// *(stub|TODO: real|placeholder|not yet)' crates byroredux tools --exclude-dir=nifskope

The first command should return nothing — any hit is notable. For each: reachable from a shipped CLI flag or smoke test →
MEDIUM; functions returning None / Vec::new() / Default::default() under a stub comment; trait impls with empty bodies
the trait docs say should do work; console commands in byroredux/src/commands/ that no-op or print "TODO".
Production-unreachable features are the yield here: a public builder/system with zero non-test callers whose docs or
ROADMAP say it is wired (a ruleset builder that main never calls; a component nothing inserts). For a candidate
name: grep -rnw name crates byroredux tools --include='*.rs' and drop test-only hits. Cross-check per-game ESM record
coverage in crates/plugin/src/esm/records/ against the ROADMAP.md compat matrix (wired vs stubbed).

Dimension 7: Magic Numbers & Hardcoded Constants

Paths: crates/nif/src/blocks/, crates/renderer/, crates/plugin/src/esm/records/
First step: read the version-gate and budget sites; do not regex blindly (most literals are legitimate)

• Bare numeric literals compared against version codes in crates/nif/src/blocks/ → a NifVersion constant.
• Vulkan MAX_* / MIN_* hardcoded inline → vk::PhysicalDeviceLimits or a named constant.
• Shader #define provenance: every define is generated from crates/renderer/src/shader_constants_data.rs
  (include!d by shader_constants.rs and crates/renderer/build.rs, which emits shaders/include/shader_constants.glsl);
  source-scan tests in shader_constants.rs police the generated header. The check is "every shader #define / loop
  bound / budget literal that bypasses that file" (lockstep risk HIGH — feedback_shader_struct_sync).
• GPU #[repr(C)] size literals: reference the layout pins (Dim 3 has the recipe and the guards); an inline size
  literal that should reference them is a code finding here, a stale prose figure is Dim 3.
• Frame/ray/cache budgets (MAX_TOTAL_BONES, MAX_MATERIALS, …) scattered vs one tunable module; ESM sub-record sizes
  hardcoded (if data.len() == 24) → a named constant on the record struct.
• Do NOT flag protocol-defined magic: FourCC tags, BSA/NIF/BA2 magic, Vulkan format enums.

Dimension 8: Dead Code & Backwards-Compat Cruft

Paths: crates/, byroredux/, tools/, every Cargo.toml
First step: cargo clippy --workspace --keep-going -- -D warnings (the CI gate) and the greps below

grep -RInE 'allow\(dead_code\)' crates byroredux tools --exclude-dir=nifskope
grep -RInE '#\[deprecated\]|// *removed:|_unused|fn .*_unused' crates byroredux
cargo machete 2>/dev/null || echo "cargo machete not installed — scan Cargo.toml deps vs use stmts"

• Unused-dependency gate: CI job Unused dependencies runs cargo machete (#3890) — a red run is a finding. Its
  blind spots are macro-only and re-export-only use (the job comment says so); the hand scan earns its keep there and on
  deps used only under a non-default feature.
• Clippy gate (CI job Test + Check + Clippy runs cargo clippy --workspace --keep-going -- -D warnings plus a
  renderer-only undocumented_unsafe_blocks re-lint, #5121; the toolchain is pinned at 1.96.0 by rust-toolchain.toml since #5308 (b24cb46b6) after floating
  stable let rustc 1.99's chunks_exact_to_as_chunks wave turn the gate red for four days — rustc 1.96 (800802516) and 1.98
  (530c9e7aa) raised lints on untouched code too; a toolchain bump is now a deliberate commit, and the distro cargo ignores
  the pin; wasmtime 49's MSRV is exactly 1.96.0, 15a6b1d2d, so the pin can rise but not fall). A red gate is a finding:
  list each failing lint and file, plus the crates it left unlinted — --keep-going lints every crate that still
  compiles, but a crate red under clippy leaves its dependents unchecked (a red byroredux-nif skips the bin).
  Run --all-targets --keep-going too; test/example-target lints are outside the CI gate — report them as a
  lower-priority bucket (the workspace example build itself went red once, the #3894 class, 530c9e7aa). Check every
  #[allow(clippy::…)] carries a reason comment (the house style for too_many_arguments).
• Each #[allow(dead_code)]: called now or still dead? Delete if dead. pub fn in a private module nobody imports (cargo +nightly rustc -p <crate> -- -W unused);
  mod.rs/lib.rs re-exports with no consumer; _-prefixed params that survived a refactor (delete, do not rename);
  // removed: … breadcrumbs (delete completely); re-exports of deleted types "for compatibility" (no external consumers
  exist — pure rot); #[deprecated] items with no consumers; Cargo.toml feature flags with one branch always on/off.
• Do NOT flag: cfg(test) / cfg(debug_assertions)-gated code, FFI boundary functions, or the public API of a
  workspace-internal crate a future binary will consume (note it rather than deleting).

Dimension 9: Test Hygiene

Paths: every #[test]/#[ignore] site, .github/workflows/ci.yml, byroredux/tests/
First step: the ignore-triage command below

# All ignores (matches BOTH `#[ignore]` and the reason form `#[ignore = "…"]` — do not tighten to a bare `]`)
grep -RInE '^[[:space:]]*#\[ignore' --include='*.rs' crates byroredux tools
# Ignores whose reason is NOT a data/hardware gate — the triage input
grep -RInE '^[[:space:]]*#\[ignore' --include='*.rs' crates byroredux tools | grep -viE 'game data|installed|\.(bsa|ba2|esm)|BSA|BA2|ESM|audio device|vulkan|gpu|rt-capable|display|opt-in|corpus|real (data|master)|steam|env'

• Guards: workspace_hygiene_tests (byroredux/src/workspace_hygiene_tests.rs) — every_ignore_attribute_carries_a_reason
  (a bare #[ignore] fails; the reason text is the triage input), no_tmp_scratch_examples_are_committed (_tmp_* /
  tmp_* probes) plus no_committed_example_self_describes_as_disposable (module doc opening "throwaway / one-off / TEMP",
  #5114), no_test_attribute_is_stacked_on_another, and the classify_pbr check (Dim 3). Confirm they are not themselves #[ignore]d.
• Most #[ignore]s gate Vulkan / game-data / audio tests — not debt. Triage the remainder: does a reason name an issue
  (gh issue view N), and if it guards a closed CRITICAL/HIGH fix → MEDIUM.
• Tests with only smoke assertions (assert!(result.is_ok())); commented-out assertions in passing tests; tests that
  println! without an assert.
• Feature gates vs CI lanes: enumerate [features] (grep -A6 '^\[features\]' crates/*/Cargo.toml byroredux/Cargo.toml)
  and compare with .github/workflows/ci.yml — non-default states run only where a lane names them (today: core
inspect/save and scripting save via workspace unification, byroredux-spt --features recon, byroredux without debug-server / with
  tracing-tracy / with dhat-heap, byroredux-nif --features dhat-heap). A feature or required-features target with no
  lane can rot silently (the class #3894 / #4387 / #4390 closed).
• byroredux/tests/golden_frames.rs (opts into --ignored) — still runnable, golden images current.
• Cross-reference "must not regress" lines in other audit skills — each named regression test still exists and is not
  #[ignore]d (_audit-validate.sh advisory names symbols that are gone).

Cross-Dimension Dedup

A TODO inside a dead function reports under Dim 8, not also Dim 5. Material doc rot reports under Dim 3, not Dim 8. A stale
GPU-size doc comment is Dim 3; a stale GPU-size code literal is Dim 7. NIFAL/material translation correctness is out of
scope (/audit-nifal) — this audit owns only the debt around that tier.

Phase 3: Merge

1. Read all /tmp/audit/tech-debt/dim_*.md.
2. Combine into docs/audits/AUDIT_TECH_DEBT_<TODAY>.md: Executive Summary (findings by severity + delta vs
   baseline.txt), Baseline Snapshot (the Phase-1 counts), Top 10 Quick Wins (trivial/small effort),
   Top 5 Medium Investments (splits, consolidations), Findings (HIGH → MEDIUM → LOW, then by dimension),
   Deferred (gated on an in-progress milestone; name it).
3. Remove cross-dimension duplicates per the rules above.

Phase 4: Cleanup

rm -rf /tmp/audit/tech-debt; tell the user the report is ready; suggest
/audit-publish docs/audits/AUDIT_TECH_DEBT_<TODAY>.md.

GitHub Labels

Findings publish under tech-debt (plus the standard <severity> and <domain> labels; /audit-publish applies it for
TECH_DEBT reports). Two sibling kind labels split the bucket — apply the one that matches:

• doc-rot — documentation drifted from code (stale ROADMAP row, a SKILL naming a deleted symbol, a comment describing
  removed behaviour); publishes as documentation, not bug.
• test-gap — missing, vacuous or non-asserting coverage (an #[ignore]d test with no data gate, an entry point with
  zero tests).
  Pure debt (dead code, duplication, magic numbers, oversized files, stale markers) stays tech-debt + bug.