nightly 2026-09-07 (v0.3.13-beta.22): it never built on Linux — a cfg! that was never compiled out, and a libc bump that broke the Python runtime - #315

What this run found

origin/master did not build on Linux. Not "had warnings", not "a test failed" — cargo build --workspace produced no binary. The dev host moved to Arch/Omarchy in September and, as far as this
run can tell, nothing had compiled the workspace there since. Verified against a pristine
origin/master before touching anything, so this is a pre-existing break, not something the
dependency sweep introduced.

Two independent causes, and the same reason neither was caught: a platform gate is only tested by
the platform it excludes.

1. A runtime cfg! where a compile-time #[cfg] was meant

crates/nanna-scripting/src/bridge.rs:746 picked the exec shell with:

let mut cmd = if cfg!(windows) { /* Git Bash / PowerShell routing */ } else { /* sh */ };

cfg!(windows) is a boolean expression, not conditional compilation. It selects the correct
branch at runtime — so this was never a behaviour bug — but both arms are still type-checked on
every platform, and the four helpers the Windows arm calls (strip_outer_quotes,
classify_windows_command, git_bash_path, WinShell) are #[cfg(windows)] and genuinely absent
on Linux. Five E0425/E0433, and nanna-scripting — hence the nanna binary — could not compile.

Fixed by splitting into #[cfg(windows)] / #[cfg(not(windows))] bindings. No behaviour change on
Windows: the same arm is selected, just at compile time. Net −7/+5 lines.

Also removed normalize_drive_paths's #[cfg(not(windows))] identity stub — its only non-test caller
is #[cfg(windows)] and its tests are #[cfg(all(test, windows))], so it was dead on every
platform and merely invisible while the crate never compiled here. Deleted rather than #[allow]-ed.

2. libc 0.2.187 broke the vendored Python runtime

libc corrected POSIX_SPAWN_SETSID from c_int to c_short on linux-gnu — correctly; glibc really
does store spawn flags in a short. But rustpython-vm 0.5.0 hands that constant straight to
nix::spawn::PosixSpawnFlags::from_bits_retain, which nix 0.30 types as c_int. E0308 at
rustpython-vm-0.5.0/src/stdlib/posix.rs:1812, killing the python feature.

The fix is already merged upstream and never released — RustPython #8343 "Fix building against
new libc", merged 2026-07-22; 0.5.0 is still
the newest on crates.io. So libc is held at 0.2.186.

That window is narrower than it looks: rustpython-stdlib 0.5.0 itself requires libc ^0.2.183, so
exactly four releases satisfy both constraints — which is why a bare cargo update lands outside it
every single time.

Increments landed

1. fix(scripting) — both Linux build fixes + the dependency freshness sweep. These could not be
   split: neither is green without the other, so the smallest green unit contains both.
2. test(storage) — a version-ceiling guard, the mirror of the existing unification guard.
3. ci — compile-tests-linux, an ubuntu-latest mirror of the Windows test-compile job.
4. chore(release) — v0.3.13-beta.22 prep.

Dependency freshness

• ocrs 0.12 → 0.13 + rten 0.24 → 0.26 — unblocks a pin that three previous runs recorded as
  "not ours to fix". ocrs 0.13.0 shipped against rten 0.26, so the pair moves together,
  rten/rten-tensor stay unified, and OcrEngineParams { detection_model, recognition_model } —
  the exact call the roadmap predicted would fail — compiles with zero source changes.
• wide 1.6 → 1.7, playwright-rs 0.16 → 0.17, deno_core 0.410 → 0.411, plus the compatible
  cargo update sweep. malachite-bigint re-pinned to 0.9.2 as documented.
• Rejected on purpose: criterion 0.8 → "0.7" (a downgrade, as the roadmap warns);
  typescript 7 (vue-tsc still cannot load it); vuedraggable left at ^4.1.0 — a blanket
  pnpm update --latest would have silently downgraded it to the Vue-2 line, so packages were
  upgraded explicitly instead.
• Frontend: @tiptap/* 3.31.3, vitest 4 → 5, @lucide/vue 1.42, vue-router 5.3.1,
  @playwright/test 1.63, happy-dom 20.14, @vue/test-utils 2.5, postcss 8.5.28, Tauri plugins
  — all with zero source changes.

The new guard

held_back_crates_stay_below_their_ceiling catches a crate resolving to one version that is too
new, where the August guard catches two versions. CeilingCrate carries a lift_when field
beside reason/remedy: a ceiling holds back upstream fixes, so recording the condition that
retires it means the pin gets deleted on purpose rather than renewed forever.

Version comparison is positional and dependency-free, and compares components numerically on
purpose — as text, "0.2.9" sorts above "0.2.186", which would make the guard silently useless. A
negative-space test pins that.

Verified it fires, not just that it was written: re-applying cargo update -p libc --precise 0.2.189 makes it report libc resolved to 0.2.189 but must stay at or below 0.2.186 with the remedy
and the lift condition. 5 tests, 0.00s.

Verification

All on Linux, after the fixes:

Real-binary check. The built Linux daemon, booted against a scratch config (NANNA_CONFIG_PATH

• --data-dir + --no-pid-file, ports 51997/51998 — never the operator's), reaches Daemon ready,
  serves IPC and health, answers GET /health with
  {"status":"ok","version":"0.3.11","uptime_secs":1}, handles SIGTERM cleanly, and logs zero
  panics. The only errors are Ollama being unreachable (not installed on this host), which the
  readiness-wait path handles as designed rather than by burning retry budget.

Benchmark (wide touches nanna-simd, so the sweep is perf-relevant):

Inside budget at every size. bench/BASELINE.md was deliberately not updated: the recorded
numbers are Windows, these are Linux, and the box was at load average 112 with a second Rust build
running — three uncontrolled variables at once. Contention can only have hurt these numbers, so
passing under it is a sound pass, but it is not a measurement worth enshrining. Filed as a re-measure.

Not verified

• The Tauri GUI on Linux. cargo build and cargo test are green; the desktop app is not part of
  that claim, and nanna-gui stays excluded from the new CI job because it needs WebKitGTK system
  packages. Filed.
• compile-tests-linux itself has never run. It has now: it passed on this PR's first CI
  run in 7m54s, with no extra system packages needed. (Kept here rather than deleted so the claim
  is traceable — it was written before the run and is now settled.)

Environment notes (need your attention)

• The dev tree was root-owned with no passwordless sudo when this run started, which blocked the
  documented worktree setup entirely. You fixed it mid-run. One leftover I fixed myself:
  /mnt/deepmem/.pnpm-store was still root-owned, failing every pnpm install with
  ERR_PNPM_EACCES; I chowned it to you.
• A stale nanna-nightly.lock from 2026-09-02 and an orphaned ../nanna-nightly worktree were
  left by a run that died early (its branch had no commits, so nothing was lost). Both cleared.
• The mummu-dev-routine ran concurrently for most of this session. Load average hit 112 with
  87% iowait, several rust-lld linkers at ~4 GB each; one cargo test was killed outright and
  cargo was SIGKILLed twice early on. Everything here was re-run to completion at -j 3/-j 4, but
  the two nightlies contending for one disk is worth scheduling around.
• There is an un-dispatched release backlog: RELEASE_NOTES.md on master describes
  v0.3.11-beta.20, PR nightly 2026-08-28 (v0.3.12-beta.21): live-and-wrong code — reactions that inverted their meaning, two unbounded maps, and read_pdf that never worked #285 prepares v0.3.12-beta.21, and the newest published tag is
  v0.3.10-beta.19.

Post-merge — cutting the release

gh workflow run release.yml -f version=0.3.13-beta.22 -f release_type=pre-release
# then, from the signed artifacts:
node gui/scripts/make-updater-manifest.mjs v0.3.13-beta.22 "Linux builds at all now: a cfg! that was never compiled out, and a libc bump that broke the Python runtime"
# and open a follow-up PR committing .updater/latest.json — that is the publish

🤖 Generated with Claude Code