"This repository exists to deposit AI garbage that actually makes Zig better and works with hyper-modular code." — Daniel Campos Ramos, founder, EchoSystems AI Studios, 2026-08-22 (with a nod to the Zig president's assessment that AI-assisted contributions are "invariably garbage" — the full story, and the full credit chain, live in
PROVENANCE.md)
Maintained by EchoSystems AI Studios (Daniel Campos Ramos) as a build-station
toolchain fork. Upstream: Zig 0.16.0, imported verbatim from
zig-0.16.0.tar.xz (sha256 43186959edc87d5c7a1be7b7d2a25efffd22ce5807c7af99067f86f99641bfdf).
Upstream's own README is preserved at README.upstream.md; upstream's MIT license
(LICENSE) governs and is preserved unmodified.
At 15:48:17.019 UTC, root operationally promoted the verified compiler
0.16.0+cgm.0399d2b19b (SHA-256 d6b4168f…) to the station default:
PROMOTED/zig now selects durable PROMOTED/stage3-d6b4168f/bin/zig.
The old stage3-046d6833 remains preserved. The
promotion receipt records
the exact binary/library identities, successful version/library discovery and
Daniel's verification-backed operational disposition. Existing functional and
control evidence is the primary basis; the two real-project attempts both
failed at matching project guards, not newly successful product builds.
Historical clean-machine cost acceptance remains UNFULFILLED, with no
retroactive performance pass. Public prerelease/latest-release metadata and
archive bytes are unchanged. Earlier candidate/default statements below retain
their historical state; this note identifies the current station default.
Building a large hyper-modular project (≈1,800 named modules / ≈2,300 files analyzed
as one compilation unit) reproducibly crashes the zig 0.16.0 frontend with a silent
SIGSEGV: a u32 0xFFFFFFFF "none" sentinel is dereferenced as a live index into an
8-byte-element table (faulting address decomposes exactly as base + 0xFFFFFFFF*8,
confirmed in two independent builds with an identical backtrace tail). The crash is
invariant across warm/cold cache, 8MiB/unlimited stack, self-hosted/LLVM backends,
and incremental on/off. Two superficially similar upstream issues were reproduced
locally and refuted by discriminator. The shipped release binary is built
ReleaseFast/stripped, so the failure is silent; a ReleaseSafe rebuild makes it name
its own site.
The Zig project's contribution policy (2026-04) does not accept content generated, edited, brainstormed, or debugged with LLM assistance — in issues, pull requests, or bug-tracker comments. This defect was diagnosed end-to-end by a human+AI partnership, so it cannot be reported or fixed upstream under that policy without misrepresenting its provenance, which we will not do. This fork is the lawful remaining channel: MIT licensing is independent of contribution policy.
Scope is deliberately minimal: a patchset on upstream 0.16.0, rebase-friendly, no language divergence. If upstream ever fixes the defect class, this patchset shrinks toward zero and the fork retires.
Historical snapshot at 63effbe9 (before the 2026-09-06 candidate): 57 commits since the verbatim tarball import
(754b7a38), touching 63 files, +12,744 / −109 lines. By directory: compiler
src/ 13 files (+3,007 / −97), lib/std/ 6 files (+575 / −2), lib/compiler/
1 file (+347 / −5), documentation 13 files (+4,071), partner_tools/ 22 files
(+4,527), .claude/ 2 files (+197). Every row below is drawn from a commit
message, a verification row, PROMOTED/RECORD.md, or docs/crown/BUILDING.md;
where the fork's record holds no measurement of a change's effect, the row says
not yet measured rather than claiming one.
The following old046d description and table labels preserve the earlier promoted inventory, not today's default pointer. Their measured facts remain unchanged; use the current station note above for the later promotion.
Which of this is in the compiler the station actually runs. The promoted
binary PROMOTED/stage3-046d6833/bin/zig (ReleaseSafe, sha256 046d6833…) was
built from the tree at b991cb16 on branch patch/005-auto-hardware-threading;
no src/ or lib/ file changed between that commit and the historical
63effbe9 tip. The historical compiler rows below therefore describe that
promoted binary. This does not include the new d3347292c9 ready-index or
0399d2b19b logging correction: their rebuilt candidates have bounded runtime
evidence but are not promoted code. The older unmeasured rows and V12 part-2
residual under group (c) also remain open. The practical result the maintainer
reports (2026-08-30, under that binary): a full build of the ≈1,800-module
workload that on stock 0.16.0 never completed — it died silently after more than
two hours, every time — now completes in under 45 minutes. That figure is a
station-side wall-clock reading, not a harness row, and it is not a speed claim
against any other configuration.
Row status: PROMOTED = in stage3-046d6833; PROMOTED, ships OFF = in the
binary but only behind an opt-in flag; TOOLING = Python/skills, not in the
binary; DOCS.
Commit d3347292c9 adds an optional indexed heap for layered ready selection:
O(1) peeks and O(log n) repairs, current-map tie ordering, rank-epoch invalidation,
and original scan fallback on optional-allocation failure or transient ownership.
Insertion stays default; this does not parallelize Sema. Root independently
reran seven helper tests (4096 mixed transitions). The logging-fixed successor
at 0399d2b19b has been rebuilt and exercised through 47 runner-accepted
incremental updates. An isolated observation build compared 10,678 indexed
choices with the original scan on the same live ready sets: all matched.
Real insertion/removal, epoch, abort/recovery, tier and cycle paths were reached.
Deliberately broken-hook controls, integration OOM and end-to-end performance
remain open; the shadow's extra scans are measurement overhead, not shipping
code. The reused runner's diagnostic-completeness limitations are explicit in
the live-shadow receipt.
The incremental runner's completeness checks were then repaired: it now refuses missing expected diagnostic tails and unchecked terminal outcomes. Three real controls that the old runner incorrectly accepted each produce the intended named failure in the repaired runner. A fresh repaired build then passed all 47 native update expectations plus three separate sema updates; the exact coverage and remaining branches are in the runner receipt. The live-shadow selection checker also has its own deliberate failure and checksum-restored rebuild/run, documented in the self-control receipt. Neither result substitutes for the still-open lost-hook and integration-OOM controls.
Update 2026-09-07: the lost ordinary append hook now has its intended named failure, exact source restoration, and a separately rebuilt successful rerun: four expected updates, 2,095 same-state agreements and two active-growing puts. See the lost-append control. Lost-update-invalidation, default-entry and integration-OOM controls remain open; this is not promotion or performance evidence.
Further update 2026-09-07: the lost-update-invalidation control now fails at a retained memo count of 98 before second-update processing, then passes all four outcomes after exact restoration and a separate rebuild, with 2,095 same-state agreements and populated memo reset checks at 98 and 3. The short-circuited predicate does not separately negative-control each index state. Default-entry and integration-OOM controls remain open; no speed or promotion claim follows.
Further default-mode update 2026-09-07: the default-entry control now has a named layered prepare-entry failure while explicit insertion and omitted order each pass all three outcomes with four zero-capacity/memo exits (including compiler_rt). Exact restoration and a separate rebuild pass layered order with 2,083 same-state agreements. The earlier prepare panic masks the append tripwire; integration OOM remains open, with no speed or promotion claim.
Superseding status 2026-09-07: the earlier open-status paragraphs remain as history. The first nodes-preparation-allocation control is now complete: 4/4 compiler builds and 5/5 intended runtime phase outcomes (four successful runs and one expected named tripwire failure), followed through exact clean restoration and recovery. The shipping candidate remains uninstrumented. Broader OOM coverage, post-abort rearming and performance remain unproved; this finite control does not establish foundation or release readiness.
The release-cost packet is materialized locally, but 0/12 timed compiles have run. Three CPU-activity holds prevented launch; neither is a failed timing sample. Package, companion evidence bundle, online main synchronization, release and promotion remain open.
Superseding executed-cost status 2026-09-07: the earlier 0/12 note is retained as history. The executed cost receipt now records 12/12 successful compiles and 13 retained non-launch holds, without retries or replacement samples. All three wall comparisons fall inside the existing heuristic, but old/default wins 3/3 paired slots and the candidate's median default RSS is +356 KiB. The clean-machine criterion remains UNFULFILLED: these are ambient-workstation-qualified observations, not performance equivalence or promotion clearance. Candidate packaging may be staged and smoke-verified separately while unpromoted; package/evidence, publication/main-sync, release and system-default gates remain open and unexecuted in this packet. No foundation or crown stage is declared complete.
The same commit corrects status-tool promotion authority, diagnostic/sabotage instructions and the documented ReleaseSafe build switches. The fresh candidate build uses safety checks, debug extensions, logging and a baseline CPU target; that target does not remove its system LLVM/runtime dependencies or prove portability. The logging correction preserves mode defaults and filters debug output by requested scope; matching/nonmatching/no-scope controls are documented in the candidate verification. Neither a new package nor promotion is claimed here. See the source review and the recipe controls.
Superseding local packaging status 2026-09-07: the unchanged candidate package is locally archived, extracted and smoke-verified, and its separate evidence companion is now assembled with 20,218/20,218 members byte-verified. Root's separate readback agrees at the shared-parser limits in the package and companion receipt. The companion contains the earlier package-verification receipt, not this later documentation successor; neither frozen archive was repacked. Later observer implementation/integration remains outside those snapshots. Publication, online main synchronization, release and default promotion remain open; the clean-machine cost criterion is still UNFULFILLED. These local packaging results do not establish performance equivalence or crown completion.
| Change | Where | Commits | Status | Observed effect |
|---|---|---|---|---|
InternPool: per-thread Index exhaustion is a named panic with remedy (thread, index reached, limit, split, -j remedy), not a silent SIGSEGV. Upstream's guard is an assert, compiled out of every ReleaseFast build. |
src/InternPool.zig |
8a3b1f1f, merged a7f31c36 (patch/001) |
PROMOTED | Fired live in production twice and printed the u30 >> 4 ceiling to the digit; the ≈1,800-module workload then completed where every stock compiler died (exit 0, 18 min 10 s, 18.4 GiB peak — a7f31c36). Neutral reproducer in the commit: 37 s, 1.3 GiB, exit 134 on the safe build. |
InternPool: CaptureValue widened from packed struct(u32){tag: u2, idx: u30} to two u32 words, so Index goes 30 → 31 bits. Bit 31 is not taken: Air.Inst.Ref owns it, and that is refused by name in both files. |
src/InternPool.zig, src/Air.zig |
8690a9c7 |
PROMOTED | Per-partition ceiling at tid_width = 4 goes 67,108,863 → 134,217,727 (the incident overflowed at 67,108,864). The dossier predicted 4×; the tree measured 2×, and the correction is recorded where the prediction was made. Also removes a latent Nav.Index truncation. H3 functional suite identical to the pre-widening control (3274/3426 steps, identical failure sets, 0 unique to this tree). |
ThreadPlan: index_bits 30 → 31, so every ceiling the report line prints doubles with the encoding instead of advertising half the real headroom. |
src/ThreadPlan.zig |
0e91a799 |
PROMOTED | V1 GREEN 4/4 — the index width is derived from the compiler's own report as 31 bits, confirming ThreadPlan.index_bits and InternPool.getIndexMask agree end to end. |
| Change | Where | Commits | Status | Observed effect |
|---|---|---|---|---|
Zcu: file exists in modules 'a' and 'b' names the double-owned file, not whichever importer happened to be scanned (.file = res.file, one field). Two test snapshots that had pinned the defect as expected output are tightened, not weakened; a fixture is added. |
src/Zcu.zig, src/Zcu/PerThread.zig, test/compile_errors.zig, test/incremental/change_module, test/fixtures/file_in_multiple_modules/ |
76a0b267 (patch/003) |
PROMOTED | H2 GREEN 4/4 expected lines verbatim; H5 GREEN — the pre-fix compiler says mod_b.zig:1:1, the fixed one dupe.zig:1:1 (the guard has been seen red). |
-femit-module-graph[=path]: the compiler writes its resolved module graph as JSON (module identity, root paths, dependency edges, graph roots, per-module cache knobs) at the instant module_roots is complete and before any analysis. Flag-gated; folded into the cache hash so a warm-cache rerun cannot silently no-op; refuses by name when there is no Zcu. |
src/main.zig, src/Compilation.zig, src/Compilation/EmitModuleGraph.zig, src/dev.zig, lib/std/zig.zig |
7ef21e28 |
PROMOTED | Stock invocations byte-identical (additive; 0 deletions). Consumed by ast-check --module-graph. Effect on a real workload: not yet measured (parse- and type-check verified; adversarially reviewed, two correctness defects fixed before landing). |
--time-report-json <path>: the same --time-report counters written once to a file after the update, with no --listen required; schema cgm-zig.time-report.v1, header only (per-decl payload and LLVM pass timings reported as counts). Refuses --listen by name instead of exiting 0 with no file. Transport only: no new counter, no new timer. |
src/main.zig |
95cea4e2, abcd7bba |
PROMOTED | Made V4 measurable at all: cpu_ns_link is 11.2× real_ns_link_flush, so the linker ran beside analysis. Before this flag the only sink was a blocking web server. |
The thread-plan report line is silent under --listen (if (listen == .none) thread_plan.report()), the test the compiler already made for --time-report. zig build and jitCmd keep theirs; they own the terminal. |
src/main.zig, src/ThreadPlan.zig |
f52308a4 |
PROMOTED | V-BR: a green zig build of a 13-step fixture went from 6 ^error: lines + 6 native failure markers (first build) to 0 / 0 (reference 0 / 0). Under --listen the fork is now byte-identical to stock on both the artifact and the diagnostic stream. |
| Change | Where | Commits | Status | Observed effect |
|---|---|---|---|---|
std.Thread.Topology: physical cores and SMT siblings, every sibling set intersected with sched_getaffinity before it is counted. physical is ?usize and null means unknown — never 1, never logical / 2; a partial probe is discarded whole. Linux via sysfs then /proc/cpuinfo; Darwin via hw.physicalcpu (read-verified only); everything else .unknown. |
lib/std/Thread/Topology.zig (new), lib/std/Thread.zig |
7106d368 (patch/005a) |
PROMOTED | V-S1a GREEN 7 of 7 masks match the sysfs oracle, including the null a mixed mask must produce; the same probe refuses to build against the reference compiler (control: it is fork-added). V14, the affinity-intersection sabotage control: not yet measured (patch prepared and git apply --check clean; the rebuild was not fired). cgroup CPU quota is named as invisible to this probe, not approximated. |
ThreadPlan: the worker count is split from the InternPool partition count. M_wide = -j<N> orelse logical; K = --intern-partitions orelse 1 << ceil_log2(max(physical orelse logical, 4)), capped by the representable partition count. AstGen runs on the wide lane: workerUpdateFile holds a thread id only for its import-discovery tail, and updateFile / lockAndClearFileCompileError / reportRetryableFileError now take *Zcu, so the rule is enforced by signature. --intern-partitions reaches every sub-compilation's pool by default rather than through twelve forwarding sites. |
src/ThreadPlan.zig (new), src/main.zig, src/Compilation.zig, src/InternPool.zig, src/Sema.zig, src/Zcu.zig, src/Zcu/PerThread.zig |
016d8987 (patch/005b) |
PROMOTED | V2-EXP GREEN: -j1 → workers 1, partitions 8, 268,435,455 items per partition — -j moves workers and leaves K alone. V5 GREEN: 1,200 files at -j64, rc=0, digest identical to -j1. On the 6c/12t reference host the derived K=8 gives 4.00× the incident's ceiling — derived from the constants; the reproduction on the ≈1,800-module workload (V7a/V7b) is not yet measured in the harness (blocked by charter). V13/V13-MM, the SMT payoff on the wide lane: INCONCLUSIVE — wall clock inside the noise floor; real_ns_files favours the wider lane by 23.6%, a note, not a claim. Named residual (R12): race freedom of the lane split is read- and type-verified only; ThreadSanitizer is unbuildable here (linux/scc.h), the Helgrind substitute was retracted for cross-build comparison, and the standing evidence is determinism — V12-P1A 1 distinct digest of 5 on the fan-out, identical to its own -j1; V12-P1B .text 1 of 5; an independent direct check 1 of 20 on both arms. |
Edges-first ordering. Build runner: --step-order=layered|random|declared (default layered: depth ASC, fan-in DESC, name ASC over the ready set; random kept as the missing-edge fuzzer). Inside one compilation: --analysis-order=insertion|layered — a priority over the ready set, never a change to the set, because eager leaf analysis would be language divergence. |
lib/compiler/build_runner.zig, src/Compilation/ModuleRanking.zig (new), src/Compilation.zig, src/Zcu.zig, src/Zcu/PerThread.zig, src/main.zig |
2a9ca530 (patch/005c) |
PROMOTED (step order); PROMOTED, ships OFF (--analysis-order=layered) |
V10 GREEN: layered 3.567 s vs random 3.631 s vs declared 3.760 s over 13 steps — a weak green, inside the noise floor against random. V8 RED for the feature: layered analysis is +6.59% slower than insertion (3 of 3 paired slots), so it stayed off; the new optional index is now authored but has no end-to-end timing result yet. The default remains insertion. V9 GREEN: a cyclic module graph does not hang the ranker. |
Rider 1: Io.Threaded.concurrent_reserve — io.async admission leaves a reserve so io.concurrent (the linker's slot) is not starved by bulk work. Default 0 makes both predicates byte-for-byte the stock arithmetic; the compiler sets 1 only when at least two async slots exist. |
lib/std/Io/Threaded.zig, src/main.zig |
0711a34e |
PROMOTED | V4 GREEN (link ran beside analysis — above). V15 GREEN: 0 of 881 thread-samples parked in Id.acquire, with 4,472 Zig frames resolved as the instrument control, so the dossier's proposed admission gate was retired by measurement. V-S2a/V-S2b sabotage controls: not yet measured (the stale contexts were repaired and apply-check verified in the 2026-09-06 correction; sabotage rebuilds remain unrun). |
Rider 2: -j reaches child compilers. std.Build.Graph.child_jobs: ?u32 (null = stock), Step.Compile appends -j<N> when set; zig build --child-jobs=keep|share|N. The default shipped as share and reverted to keep when its own gate fired; share stays a selectable member. |
lib/compiler/build_runner.zig, lib/std/Build.zig, lib/std/Build/Step/Compile.zig |
9a7cbab8, 50a7b87f |
PROMOTED (default keep); share ships OFF |
V-S4a GREEN: peak worker threads across a -j4 build on an 8-CPU mask, keep 35 → share 10. V-S4b RED for share: +6.55% slower, 3 of 3 paired slots; peak process-tree RSS share 1,065,244 KB vs keep 1,104,468 KB. Caveat named in the code: a 13-step fixture of short compiles is the regime where the cure costs most; a large project could invert it. |
ThreadPlan: a starved thread-id pool is impossible to derive (min_derived_basis = 4) and refused by name when an explicit --intern-partitions would leave 0 allocating lanes. This was the fork's own regression — the (K, M_wide) split made the state reachable — found by the harness and fixed inside the same run. |
src/ThreadPlan.zig, src/Zcu.zig, src/main.zig |
3669ffc7 |
PROMOTED | V16 GREEN, 0 hangs of 6 configurations: build-obj hello.zig with no flags on a 2-physical/4-logical mask went from rc=124 (hung, first build) to rc=0; -j4 --intern-partitions=2 → refused, rc=1, with the remedy; -j1 --intern-partitions=2 still rc=0 (the over-fire guard). |
| Change | Where | Commits | Status | Observed effect |
|---|---|---|---|---|
zig ast-check --module-graph=<path> file.zig: after the normal check, validates the file's @import operands against a stage-0 module-graph JSON — relative .zig/.zon imports must exist on disk, named modules must appear as an import edge — and ends with one reconcilable summary (seen N: validated / exempt / skipped-unchecked). Malformed graph, stdin, or ZON mode is a named fatal. |
src/AstCheckImports.zig (new), src/main.zig |
885fada0 |
PROMOTED | Flagless ast-check byte-identical (additive; 0 deletions). Effect on a workload: not yet measured (parse-verified; three review defects fixed before landing, including a .zon false positive). |
zig ast-check f1.zig f2.zig …: batch mode over N files, skips named per file, summary R = C + F + S always; --json emits one machine-readable document with per-file status and the import statistics above. Directory arguments refuse as IsDir by name. |
src/AstCheckBatch.zig (new), src/AstCheckImports.zig, src/main.zig |
3d37ca1e |
PROMOTED | Single-file, no---json path byte-identical to upstream. Batch memory flat: 400 files at 7.6 MB with a per-file arena, where the first draft grew ≈475 KB per file (the commit's own receipt). Beyond that: not yet measured. |
| Change | Where | Commits | Status | Observed effect |
|---|---|---|---|---|
Compilation.resolveEmitPath becomes pub. Upstream never marked it; EmitModuleGraph.zig calls it cross-file. It lives in the shared history as e0bcdab2. The old local-main cherry-pick 63778997 is preserved at backup/main-before-sync-2026-09-06; local main was reconciled to online 63effbe9 before the new candidate. It retires the day upstream marks it pub. |
src/Compilation.zig |
e0bcdab2 |
PROMOTED | main had not compiled since 2026-08-22 ('resolveEmitPath' is not marked 'pub') and nobody met it because the promoted binary predated the break by sixteen minutes. After: type-check 0 errors, stage3 ninja exit 0. |
| Change | Where | Commits | Status | Observed effect |
|---|---|---|---|---|
| The fork's own record: provenance, the contribution contract, this front door. | PROVENANCE.md, CONTRIBUTING-AI.md, README.md |
7b77211c, 5ec07fe0, cc383ab8, 748540f8, 1a0d5910, b018bcac, 16f5299f |
DOCS | — |
docs/crown/: the staged plan, seven design principles, a cited internals map, the build recipe with every gotcha and its negative control, the patch/005 dossier, its verification list, and two run records including the pre-committed promotion rule. |
docs/crown/*.md |
4c76c861, 5100c24f, a5f9e0b9, da0968c5, 40a58902, 44e391fb, 9d86c443, 24155d14, 430d5077, 5d3f2d5d, 46763cb8, 2d70958a, ed863a7d, 178ba0a5, a0376dbd, 573c55e3 |
DOCS | BUILDING.md: a clean stage3 under an 8-CPU mask on a contended host is ≈41 min (the "8–9 min" figure is an idle unconstrained host); the ZIG_LIBC remedy is negative-controlled on the unpatched compiler (rc=1 → rc=0); ThreadSanitizer does not build here and a header shim is refused by name. |
partner_tools/: status eyes, one-command reproductions, oracle helpers, a patch ledger derived from git history, and helgrind_diff.py — the named TSan substitute, with its noise floor measured before it was trusted. |
partner_tools/*.py, partner_tools/README.md, .gitignore |
885fada0 (swept in), f8a661cb, 15b53145, efcbb899 |
TOOLING | Helgrind's absolute count fails its own negative control (149,968 errors on the unpatched compiler for a trivial build-obj -j2); the difference-of-signatures instrument has a measured floor of 0–7 filtered contexts across 6 same-binary pairings — and was then retracted for cross-build comparison (≈276 "new" contexts of symbolic drift against a floor of 4). |
partner_tools/vharness/: the verification harness — 38 self-registering rows (37 at its README's writing, plus V16), 7 generated fixtures, 4 reviewed sabotage patches, one driver emitting a markdown table and a JSON sidecar. No file holds a list of rows. |
partner_tools/vharness/** |
4a963cbb, e8b8f7d1, 18fc3f7e, d76d75ee, 64befd76, 3689fe3b, 56771bec, 2e591d15, b991cb16, a0376dbd |
TOOLING | Run 2 (2026-08-23, 8-CPU mask, contended): 20 GREEN · 4 RED · 9 UNKNOWN · 4 INCONCLUSIVE of 37 rows run; after three harness defects were fixed without relaxing a row and the desk's disposition, the final table in PROMOTED/RECORD.md reads 22 GREEN · 2 RED (both feature-comparison rows whose defaults already reverted, plus a pre-existing H3 identical to baseline) · 9 UNKNOWN · 4 INCONCLUSIVE. The harness found the V16 hang and the V-BR channel break; neither was a queued row. |
| Claude Code skills: the fork's working knowledge for AI partners, and a gated stub for the eventual stable-0.17 upgrade that refuses by name until upstream ships. | .claude/skills/cgm-zig/, .claude/skills/cgm-zig-release-upgrade/ |
09adf385, f70dc33a |
TOOLING | — |
The language: no syntax, semantics, or analysis-set change anywhere — edges-first is
a priority over what would be analysed anyway, never an addition to it. The
standard-library API surface, except the three additions named above
(std.Thread.Topology, Io.Threaded.concurrent_reserve with a stock-preserving
default of 0, std.Build.Graph.child_jobs with a stock-preserving default of
null) and a five-line lib/std/zig.zig enum member for the new emit kind. Every
new compiler behaviour sits behind a new flag; a stock invocation behaves stock,
and where it does not (the report line under --listen) the record above shows
it was measured and fixed. The license: upstream's MIT LICENSE is preserved
unmodified and governs.
Version policy. 0.16.0 is the base and the patch target. 0.17 enters only after upstream promotes it to stable; the 0.17-dev line was measured to fail on the motivating workload from language churn alone before reaching the crash site. Upstream Zig now lives at codeberg.org/ziglang/zig; the GitHub mirror this fork was compared against is frozen at 2025-11-26 ("README: migrated to codeberg"), and as of 2026-08-30 codeberg's newest tag is 0.16.0 — no 0.16.1, no 0.17 tag exists.
Contribution stance. This fork is not offered upstream, and nothing here is
filed there. Upstream's policy on AI-assisted contributions, and why it makes this
fork the lawful remaining channel, is described in PROVENANCE.md.
Diagnosis and patches are the joint work of Daniel Campos Ramos and AI partners including Anthropic Claude models via Claude Code and OpenAI Codex partners, with GPT 6 Astra credited for the continuing orchestration and integration. The original diagnosis and authorship remain credited to their contributors; see the full credit chain. Each patch commit carries its evidence trail and its verification limits.
This fork is deliberately AI-friendly — the inversion is the point. Start with
CONTRIBUTING-AI.md (the contract: contributions are
judged by their receipts, never by their author's species), load the working
skill at .claude/skills/cgm-zig/ (laws,
doc map, proven workflows — repro runs, oracles, bisection, the safe-compiler
build recipe), and use the stdlib-only tooling in
partner_tools/ — status eyes, one-command
reproductions, oracle helpers, and a patch ledger derived from git history
rather than hand-maintained. A gated skill for the eventual stable-0.17
upgrade sleeps at .claude/skills/cgm-zig-release-upgrade/ and refuses by
name until upstream ships.
"Tenha fé, porque até no lixão nasce flor." — Mano Brown, Racionais MC's, Vida Loka Pt. 1 "Have faith — because even in the lixão, a flower is born." (lixão: the vast open-air garbage mountain — see the translator's note in
PROVENANCE.md)
The earlier publication-open statements are preserved as history. Root published
the unpromoted prerelease 0.16.0+cgm.0399d2b19b
at 14:18:54 UTC with the two frozen archives and SHA256SUMS: its recorded API
verification matches 3/3 asset names, sizes and digests, not downloaded bytes.
The publication successor
records tag 767784fec9 and online main cf28537cb4 at that observation.
Stable 0.16.0+cgm.046d6833 remains latest; PROMOTED/zig still selects
stage3-046d6833. Clean-machine cost acceptance and default promotion remain
UNFULFILLED. Neither stopped cost series resumed, and neither frozen archive
contains the later observer/stopped-series evidence or these successor notes.
Publication establishes no speedup, performance equivalence or Crown completion.