# Refactoring backlog — OPEN items only, grouped by EXECUTION CONTEXT Lean working document. Every item here is OPEN. Completed/obsolete items and the observation journal live in the archives: [`…-archive-2026-08-10.md`](refactoring-backlog-archive-2026-08-10.md), [`…-archive-2026-08-27.md`](refactoring-backlog-archive-2026-08-27.md), [`…-archive-2026-08-30.md`](refactoring-backlog-archive-2026-08-30.md), [`…-archive-2026-08-31.md`](refactoring-backlog-archive-2026-08-31.md), [`…-archive-2026-09-02.md`](refactoring-backlog-archive-2026-09-02.md), [`…-archive-2026-09-03.md`](refactoring-backlog-archive-2026-09-03.md), [`…-archive-2026-09-07.md`](refactoring-backlog-archive-2026-09-07.md), [`…-archive-2026-09-09.md`](refactoring-backlog-archive-2026-09-09.md), [`…-archive-2026-09-17.md`](refactoring-backlog-archive-2026-09-17.md). This file shows OPEN work only + CHANGELOG.md + memory — do not resurrect without re-verifying. Legend — effort: S(mall)/M(edium)/L(arge); impact: ★ … ★★★. Prefix glossary (only the prefixes this OPEN file still uses): **F#**=the size and duplication registers. Image gaps (**CON7** onwards) live in the root [`BACKLOG.md`](../BACKLOG.md). Everything else is archive-only: **CON1–CON6** closed 2026-09-17/18 (the consumer issues, the last one measured and aligned at Flutter 3.47.4), **EX** closed on 2026-09-07/09 (the extent gates scan `linux/llm-stack` and froze its rows), **VK/AS** closed on 2026-09-17, **CC/CL/CS/AB/R#/YB/DISK/APP** on 2026-09-07, **HT/GH** before them, **QW/TC/SMK** in the 2026-09-04 waves, and the rest long before that. Last groomed: **2026-09-18, after CON4's Windows half was measured, the dependency wave landed and the two owner questions were answered** — every closed narrative moved to [`…-archive-2026-09-17.md`](refactoring-backlog-archive-2026-09-17.md) and the CHANGELOG. What stays here is **two registers**. Every earlier grooming's warning still applies: **re-derive; do not trust a number here, including these.** ## OPEN ### F1. The extent queues — what is left after every row got a verdict [M each] **`function-size.allow` and `code-complexity.allow` are the authority — do not transcribe them here.** Both are fully reviewed: **24** function rows over 80 lines and **59** `cc` rows over 15 (plus 1 nesting), every one carrying a verdict that says what its number IS. Read the reasons, not the numbers — and re-derive the counts from `verify_code_size.py` / `verify_code_complexity.py`, never from this line. **Four measurement facts that decided most of the remaining verdicts**, and that a future reader should not re-discover: * **Heredoc payloads are not shell.** `assert_pinned_versions` is 44 lines of shell around a **312**-line embedded Python program — top of the size queue and the WORST candidate on it, because splitting the shell moves 44 lines and its `cc` is **7**. Same shape: `assert_app_venv_parity` (20 around 72), `_gst_xpy_write_config` (14 around 70), `ensure_meson_cross_file` (56 around a 37-line Meson-ini template). * **Much of the `cc` here is a TABLE, not tangle**: flag and subcommand parsers (`parse_tvm_args` 13 options, `append_tvm_cmake_args` 15, `setup-dependencies.sh` `main` 5 valued flags × 11 commands), feature tables (`_ffmpeg_probe_core_codecs` is fifteen `if probe; then --enable-` lines and nothing else), and two rows where the metric is simply literal — `dump_debug_info` (cc 23) holds one decision (the `HOST_PYTHON` guard) and otherwise ~20 which-then-`--version` pairs each swallowed, and `_torch_run_setup_py` counts the size of torch's build environment. * **Refusal matrices cost safety when flattened**: `_chain_prune_archived_logs` (every branch is a refusal to delete the wrong thing), `_manifest_wrapper_gate` (the cell that decides whether a manifest would MIX releases), `install_target_packages`, `override_soundtouch_codeberg_checksum`. * **Precedence ladders where the ORDER is the contract**: `host_python_bin`, `install_abseil_headers`' five download/extract rungs, `_detect_gcc_cxxabi_header`, `configure_opencv_build_env`'s gstreamer-libdir search (the RV1-GST-PC ladder). **One row still carries a KEEP verdict worth keeping:** `media_common_init` (29) — a module loader whose load ORDER is load-bearing. A table plus a loop would read shorter and say less; the ordering is the knowledge. **The last outside-the-closure row closed on 2026-09-18.** `bump_versions.py main` (160) is now 32: `linux/scripts/tests/test-bump-versions.sh` — 40 assertions over fake specs, in-process — was the coverage it was blocked on, then the split into `_parse_args` / `_lookup` / `_sweep` / `_safe_row` / `_report_row` / `_write_phase` and the reporting helpers. Its `function-size.allow` and `code-complexity.allow` rows were deleted rather than re-baselined. Every named row left in the two registers is inside the build closure. **Two rows were given a "do not do the obvious thing" verdict; one is a row today.** `verify_comment_size.blocks` (nesting 6) is not in `code-complexity.allow`: `_walk_scan` holds the walk and `blocks` measures cc 9, nesting 4 (2026-09-25; the one nesting row is `smoke-android.sh check_ndk`, 7). Its verdict still stands for the walk: importing `verify_code_size.scan` would NARROW the scan by `SKIP_DIRS 'patches'` (the gate's own `SCAN` comment says why), which changes the gate's scope and needs a fresh `comment-size.allow` baseline. And `verify_package_names.load_arch` (17): every branch is a way the gate must not produce a FALSE verdict, the all-or-nothing partial-fetch refusal above all. ### F2. Files over ~800 lines [L each, low priority] **`file-size.allow` is the authority — do not transcribe it here.** The gate prints `files: 13 over 800 lines; 13 frozen`; read it there. All rows were reviewed and all but the one historical split are NOT split targets, each with a reason a stranger can act on. **Two verdicts worth not re-litigating.** `build-app-wheelhouse.sh` is the near-miss: the stage suites extract blocks from it **by line range**, so a file split silently re-aims them. And `smoke-runtime-image.sh` — which every earlier version of this entry nominated as THE one to split — is an explicit **NO**: its 99 functions (2026-09-25; 91 on 2026-09-18) are the `check_*` / `_probe_*` assertions and the probe-and-verdict layer they share, all over one image through one `_rt_run` under one `main()`. Its length is the number of assertions it makes about the shipped bytes, and that number growing is the gate succeeding. **`lib/app-packaging.sh` crossed 800 on 2026-09-17** (CON2's scope probe + CON6's finish-args block, 797 → 851); the row carries the NOT-a-split reason and the seam that kept it that way. Record in [`…-archive-2026-09-17.md`](refactoring-backlog-archive-2026-09-17.md). **Standing context, not a block:** `git push` is the agent's (2026-09-06) via `gh auth setup-git` + HTTPS remotes, and twelve of the fifteen `*.sh` files in `linux/scripts/lib/` source a sibling `lib/log-bootstrap.sh` (2026-09-25) — a consumer that copied ONE `lib/*.sh` file out on its own breaks on its next CI run, not here.