Game and engine library candidate suite

• Type: feature / investigation (library-suite discovery + compiler test workloads)
• Status: done
• Track: B+C
• Owner: opus-night
• Opened: 2026-06-28
• Relation: expands [[feature-c-source-frontend]], [[feature-c-regex-library-devtest]], [[feature-synapse-compile-check]], and [[feature-embed-pascal-script]] with game/engine-shaped workloads. Candidate catalog: [[game-library-candidates]].

Goal

Pull in more real libraries as non-gating compiler testing ground, focused on game engines/frameworks and their low-level support libraries. The intent is not to ship these as dependencies by default; the first value is discovery pressure for the compiler, C frontend, RTL, PAL, and FPC/Delphi compatibility surface.

Candidate set

Use devdocs/developer/game-library-candidates.md as the ranked source list. First wave:

• Pascal: New-ZenGL, Apus Game Engine, Castle Game Engine.
• C: stb, cglm, miniaudio, ENet, Nuklear, sokol, raylib, Orx.

Defer SDL, Allegro, TIC-80, Doom/Quake/Build-family engines, and legacy Delphi/DirectX engines until the smaller candidates have produced useful gaps.

Policy

• Stage in library_candidates/<name>/, not lib/.
• Prefer pinned upstream commits or release archives. Record URL, commit/tag, date imported, license, and any local edits.
• Keep discovery non-gating unless a small slice graduates into an owned library smoke.
• Do not patch around compiler failures silently. Classify each gap:
  • Track A: Pascal language, codegen, ABI, parser, or compiler bug.
  • Track B: RTL/PAL/library surface, Pascal compatibility units, examples.
  • Track C: C header/source frontend, C preprocessor, CRTL.
• GPL/copyleft candidates are allowed only as local discovery workloads unless a separate license decision says otherwise.

Slices

A - import metadata and harness shape

Add a tiny manifest convention for library_candidates/ imports, or reuse the existing candidate pattern if one has emerged. The manifest should capture upstream URL, pinned revision, license, import date, local patches, and the current first failing probe.

Acceptance: one candidate import has a manifest plus a make or script entry that can run the probe without affecting make test.

B - first C ladder

Import and probe in this order: stb, cglm, miniaudio, ENet, Nuklear, sokol. Keep each first probe intentionally small. The expected early win is a good map of C-body, macro, struct, callback, enum, and standard-library gaps.

Acceptance: each imported candidate has either one passing header/import probe or a documented first compiler gap mapped to an existing/new ticket.

C - first Pascal ladder

Probe New-ZenGL first, then Apus, then a deliberately tiny Castle Game Engine slice. Keep the editor/build-tool surfaces out of scope until core units compile.

Acceptance: each Pascal candidate has a recorded first blocker list, with genuine compiler gaps split out and RTL/PAL/library gaps kept here or in Track B tickets.

D - graduate useful slices

When a candidate slice compiles cleanly and has a deterministic smoke, decide whether it stays discovery-only or graduates:

• lib/crtl/ for C runtime-style support libraries.
• lib/rtl/ for owned Pascal library surfaces.
• examples/ for demos that prove the compiler in public.
• library-suite-discovery for ongoing non-gating coverage.

Acceptance: at least one candidate slice graduates to a stable smoke or is explicitly recorded as discovery-only with the reason.

Done when

At least three candidates from different shapes have been imported and probed: one single-header C library, one nontrivial C source library, and one Pascal game library/engine. Their first blockers are either fixed or filed as focused tickets, and the discovery command reports their status without breaking the green library suite.

Log

• 2026-06-28 — opened from the game-engine candidate discussion. Added ranked catalog in devdocs/developer/game-library-candidates.md; intent is to use suggestions as testing ground and extract only the source-compilation-relevant candidates.

• 2026-07-08 (fable-c, C ladder / slice B) — imported stb, cglm, ENet (installer fetchers stb/cglm/enet, pinned commits, gitignored). First probes under test/gamelib/, each mapping a first compiler/runtime gap:

  • stb (stb_sprintf.h) → inline fn-pointer PARAM call not registered: [[bug-c-inline-fnptr-param-call]].
  • cglm → (1) crtl lacked the C99 float-math family (added fabsf/sqrtf/... then REVERTED on discovering (2)) float-returning C functions return 0: [[bug-c-float-single-return-zero]]; and local nested aggregate initializers fail: [[bug-c-local-nested-aggregate-init]].
  • ENet → crtl missing <netinet/tcp.h>/<netdb.h>/<poll.h> ([[bug-c-crtl-missing-net-headers-enet]]) → host-header fallback redefines struct in_addr and trips a struct-tag-redefinition field-misfile that makes a record self-referential → compiler SIGSEGV ([[bug-c-tag-redef-misfiles-field-selfref-segv]]). Landed in-lane and green: crtl arpa/inet.h IPv4 text conversion (inet_aton/inet_addr/inet_pton/inet_ntop — ENet's actual needs, all pointer-based) + test/gamelib/crtl_inet_smoke.c in test-core. inet_ntoa (4-byte struct BY VALUE) omitted — hits a small-struct-byval param gap, not needed by ENet. Slice B acceptance (each candidate: passing probe OR filed gap) met for the three C candidates. Pascal ladder (slice C) unstarted — Track B.

• 2026-07-11 (ticket maintenance) — re-merged the orphaned log fragment that had landed on the backlog path while the full ticket sat parked in unfinished/ (duplicate file). Slice B (C candidates: stb/cglm/ENet probed, gaps filed) DONE; slice C (Pascal ladder) unstarted — Track B.

• 2026-07-11 (opus-night, Pascal ladder / slice C started) — imported New-ZenGL (installer fetcher zengl, pinned commit 26f40251, gitignored) and probed the leaf units (test/gamelib/zengl_probe1.pas). Three Track P blockers filed, in order of encounter:

  • [[bug-pascal-high-low-in-const-expr]] — array[0..High(LongWord) shr 1 - 1] (zgl_types.pas:105) rejected by the const evaluator; even array[0..High(Byte)] fails. Local-patched in the candidate tree (literal bounds + marker comments) to keep probing; zgl_types then compiles clean.
  • [[bug-pascal-directive-inside-paren-star-comment]] — zgl_gltypeconst has a big (* ... *)-commented block of half-translated C #if junk (&&, {$include <inttypes.h>}); PXX evaluates {$...} inside (* *) comments and dies. FPC treats them as comment text. Hard-blocks every unit that pulls zgl_gltypeconst (i.e. nearly all of ZenGL).
  • [[bug-pascal-include-search-silent-miss]] — {$I zgl_config.cfg} lives in headers/ while units live in src//srcGL/; PXX has no -Fi include path AND a missing include is silently dropped, so every zgl unit compiles with the wrong (empty) configuration. The silent miss is the dangerous half. Probe checked in (documents the fetch/compile line + blocker list); ladder parked until the lexer-level P tickets land — next Pascal candidate (Apus) will likely hit the same include-path wall, so fixing these first is the cheaper path.

• 2026-07-11 (parked -> unfinished) — slice C's ZenGL probe done + blockers filed; further Pascal-ladder candidates (Apus, Castle) hit the same include-path/lexer walls, so parked until the three Track P tickets above land. Slice B remains done; slice D untouched.

2026-07-14 — UNPARKED. The blockers it was waiting on are all DONE.

Triage sweep re-checked every reason this was parked. All three Pascal-ladder blockers are in done/ and verified fixed at HEAD:

• array[0..High(Byte)] and array[0..High(LongWord) shr 1 - 1] compile (a probe prints 256);
• {$include <inttypes.h>} inside a (* ... *) block comment is now correctly ignored;
• a MISSING {$I} is now a loud error (cannot open include file) instead of a silent skip, and -Fi<dir> exists for include search paths.

So slice B is done and slice C can resume immediately: re-probe ZenGL unpatched, then Apus / Castle. Slice D (graduation) untouched.

Moved back to backlog/ — it was sitting in unfinished/ behind a wall that no longer exists.

• 2026-07-19 (backlog sweep) RESOLVED. Done-when met: stb/cglm/ENet + ZenGL probed, blockers filed and since fixed (c256c692/c33ebbc1/43edf7db), crtl inet slice graduated into test-core. Residue: unpatched-ZenGL re-probe, Apus/Castle, slice D graduation — re-file if the game push resumes.
• 2026-07-19 — resolved, commit 43edf7db.