Experiment failure triage

These are follow-up observations from The first Filnix inventory (eaaa75f8-2149-452d-8c0e-e76d6c584029), whose recipes are frozen at b14a53e. The campaign continues with those recipes. Follow-up fixes below do not rewrite its failures or claim that blocked downstream packages now pass.

Verified fixes — 2026-09-13

pycparser 2.22: supply the preprocessor used by its tests

The original Fil-C Python package ran 130 tests and reported four failures and four errors because cpp was absent. The compiler wrapper exposes Clang but has no cpp alias. All eight failures involved preprocessing fixtures/examples.

The Python port now adds the native compiler's tools to PATH in preCheck. Only the fixture preprocessor is native; the test hook still invokes Fil-C Python 3.12.5. No tests were disabled. The full 130 tests pass.

• Original: /nix/store/b8x301fqxfkd7s0ga0yj5hdmkaq3ysz1-python3.12-pycparser-2.22-x86_64-unknown-linux-gnufilc0.drv
• Fixed: /nix/store/5slg9j9c2fp4bqhszf9zkd94jabjaw4d-python3.12-pycparser-2.22-x86_64-unknown-linux-gnufilc0.drv
• Original attempt: 509b1c36-3267-49b1-99fa-0fe50dec0d4d

The failed dependency appears under Fabric. Fixing pycparser alone does not establish that Fabric works: its recorded graph also has other failures.

QuadProg++: declare the library's math dependency

QuadProgpp compiled its static library but failed to link its example with unresolved pizlonated_sqrt and pizlonated_sqrtf. The CMake target now declares m as a PUBLIC dependency. This fixes both the example and the installed target's link interface for downstream consumers.

The package builds. A separate Fil-C C++ consumer imports the installed CMake target, links without manually adding -lm, and verifies the unique solution (1, 2) with objective value 12 for a constrained quadratic problem. CTest passes. This is a focused consumer check, not an upstream test-suite claim.

• Original: /nix/store/1cxlf9p26ynvm4cqp40b8g76fzimbq5i-quadprogpp-x86_64-unknown-linux-gnufilc0-unstable-2023-01-20.drv
• Fixed: /nix/store/hh785wgvq8p10x8s2c2pnzcy2pzwq0gm-quadprogpp-x86_64-unknown-linux-gnufilc0-unstable-2023-01-20.drv
• Consumer output: /nix/store/3hykddwyxmmizlh20q0kx2x60i9jz4l8-quadprogpp-consumer-check-x86_64-unknown-linux-gnufilc0-1
• Original attempt: 509b1c36-3267-49b1-99fa-0fe50dec0d4d

For each fixed package, comparing recursive store inputs against the failed recipe found exactly one new derivation: the package itself. The existing compiler, runtime, and other dependency derivations were reused.

Reproduce from the revision containing these fixes:

nix build .#legacyPackages.x86_64-linux.pkgsFilc.python312Packages.pycparser \
  --no-link --max-jobs 1 --cores 2 --timeout 600 -L
nix build --impure --file tests/quadprogpp.nix \
  --no-link --max-jobs 1 --cores 2 --timeout 600 -L

Build logs are retrievable with nix log on these derivations. The original attempt logs remain in the experiment state. Additional local captures and closure comparisons are in ignored results/experiment-triage-20260913/.

Deferred findings from the first batches

Initial targeted GTK follow-up, 2026-09-14

Plan 83b6d44e-5ce8-483c-80d6-5bffbdc047a4 evaluated only gobject-introspection-unwrapped, gnutls, gtk3 and gtk4 from committed revision 05ca9018d024e526347c98a29da09cd0afafeafe. All four evaluations succeeded. Introspection and GnuTLS entered the ready queue; GTK3 and GTK4 remain held by previously recorded dependency failures. The campaign's default source and 4,216 other evaluation failures were verified unchanged.

GTK3's eight blockers all enter through at-spi2-core and its systemd dependency (including kexec-tools, libapparmor, cryptsetup and libseccomp). GTK4's 32 blockers enter through gst-plugins-base and gst-plugins-bad, including the existing ALSA, Opus, codec and systemd failures. These are dependency results, not evidence that the patched GTK sources failed to compile.

The next bounded investigation is the optional dependency configuration: AT-SPI exposes systemdSupport, and GTK4's media backend pulls in the large GStreamer plugin closure. Any such recipe changes need a fresh recorded revision. This queue submission did not clear shared failures or disable tests. Full explaining chains were saved in results/gtk-gnutls-ports/{gtk3,gtk4}-queue-blockers.json.

Shared GTK/TLS repairs, 2026-09-14

The subsequent investigation fixes the scanner/tool ABI boundary across the Fil-C package scope, builds upstream-patched AT-SPI and PyGObject, and repairs DConf and Duktape failures found through GTK's shared inputs. AT-SPI now links the systemd libraries without the full tool closure; GTK4 omits its optional GStreamer video backend. Both GTK versions build and pass the Broadway consumer check; the original dependency blockers are absent from the replacement toolkit graphs. See gtk-ports.md for version choices, limitations and reproducible consumer checks.

Observed checks include 60 introspection tests, 14 DConf tests, six libproxy tests, six GLib networking test groups, and a PyGObject runtime probe. These results belong to their exact new derivations; they do not erase the original failed attempts. Separate TLS and GTK consumer checks are exposed by the flake.

Runner 0.12.2 accepts explicit source revisions when replanning queued, failed, blocked or inconclusive candidates as well as evaluation failures. The new plan records each candidate's previous recipe/result before replacing it; active and already successful candidates cannot be replaced. This allows a bounded downstream retry without changing the campaign's frozen default source or automatically reopening every evaluation failure. The regression suite has 105 passing tests.

Plan ef458359-6ce6-45f3-acb7-083c7849494f evaluates 62 selected downstream roots from 921eb54aa6366d1c71cd1cd10fcaad491e7671fd: all 62 evaluations succeed, 30 recipes enter the ready queue, and 32 have other recorded dependency failures. This is a follow-up result, not a reclassification of the original campaign. GTK3 subsequently drops its disabled devdoc output. Plan a5bc20ac-fe95-42a7-802d-706a7b06c0fb updates its 21 selected consumers from 58f7bc9f8a3b8cae3339f07c903a08f7c489c1fc, before their builds start. Plan e4c3ca39-ba4f-43e2-95c1-5ca1ecd0f434 adds 14 other selected GnuTLS evaluation failures from that same revision: apcupsd, apt, apt-cacher-ng, atop, bitlbee, chrony, conntrack-tools, corosync, coturn, dnsdist, gnupg, mailutils, rdesktop and squid. This brings the explicit follow-up selection to 76 roots.

The final separate GnuTLS check passes 526 tests with 60 skips and no failures across its TLS, certificate and slow suites. Libmicrohttpd 1.0.1 builds without additional package changes. These local checks and builds reuse the unchanged Fil-C compiler, GLib and Python derivations.

Two additional local probes identify further porting work. GtkSourceView 4.8.4 sets GLIB_VERSION_MAX_ALLOWED below 2.80, selecting the still-integer fallback of GLib's _g_type_once_init_type macros. Libsoup 2.74.3 generates enum registration code using gsize and integer once initialization; this blocks GSSDP as well. These failures are retained in results/gtk-unlock/downstream-1.log. A future GLib header correction should cover the old API branch without changing the pointer ABI; libsoup also needs its enum template adapted. The working GTK consumer checks do not imply these downstream packages pass.

Earlier findings

These are diagnostic leads, not fixes or complete root-cause analyses. Use the named original attempt to find its recorded derivations and logs.

Some candidates also stop at evaluation because Nixpkgs marks dependencies broken (for example GTK, GnuTLS, and GObject Introspection), or rejects an unfree package under the current policy. Those are evaluation refusals, not evidence that Fil-C compilation failed. The original inventory remains intact.

Scope correction — 2026-09-14

The hardened kernel variants escaped the original inventory's directory filter because their metadata position points to pkgs/top-level/linux-kernels.nix. They are excluded from the userspace experiment, including two variants whose outputs were already realized; those outputs do not establish Fil-C kernel compatibility. The inventory and runner now recognize kernel builders explicitly. Kernel headers and userspace programs stay eligible.

Batch 02a8d98c-c9bf-4961-a74a-b4456c9ed61c was cancelled while Linux 6.1.141 and 6.12.43 were compiling. Its history records the scope exclusion and preserves the worker's cancellation result and raw logs. Unrelated interrupted roots from that batch are explicitly requeued. Batch 81ce27c5-fd6d-45b8-a870-56eba8938dc8 continued running independently. Local verification is retained in results/experiment-kernel-exclusions/.