Release-candidate notes — 82936f6 (the core closure, N101–N108)

Frozen evidence for one tested commit. This file records what was verified at 82936f6d857eca7e3620793655a83b39458652a9 and is not updated as the project moves. Current implementation truth is ../capability-matrix.md; what is not done is ../limitations.md. The method is architecture.md §7.109.

Nothing has been tagged, pushed or published. This is a candidate. Making it a release is a person’s decision, and the procedure for it is at the end of this file.

What this candidate is

Nazm after the core-closure programme, N101–N108, on top of cf664fd.md’s post-v1 candidate and the N76–N100 programme: every mutant accounted for with no silent tier (N101); the compiler written in Nazm at full parity, 21 of 21 probes (N102); re-exports with one identity (N103); authority as a parameter, the inheritance bridge removed (N104); one instruction-level LIR both native backends translate, with a validator and an interpreter as oracle (N105); the M:N task pool as the default runtime, and the two performance regressions since N75 attributed (N106); and the core’s scope audited so that VERIFIED means the promise (N107). N108 adds no feature: it is this gate, and the defects it found are fixed below.

Identities

Classification (N108, §7.109)

Targets, and what was run on each

All fifteen platform runs passed at this commit.

The gate

Contained (Docker, offline, nazm-contained:1.98.1), 2026-10-05/06, serial, from clean worktrees. The full catalogue ran on the candidate the audit began from (fc509ce, N107’s records); every defect the gate found was fixed in a commit of its own, and each fix’s reach was gated again on the commit it made, the last being this one.

82936f6 changes debug builds only (an ordinary build’s code is byte-for-byte unchanged), and 230324c and b66f50f change tests and the catalogue only, so fuzzing, the benchmark and the lifecycle suite, run at 230324c, stand for this commit.

What the gate found, every item fixed before the candidate (commit, then what):

• 82936f6 — a debugger stopped on a function’s first line, not its first statement, since N105: the live gdb test (ignored in the suite: it needs the debug image) failed. The prologue’s slot zeroing had become an llvm.memset that carried the function’s line, and a guard’s continuation block took whatever line the text showed last. An intrinsic in the prologue now carries no line, and a block restates its statement’s position; two debugger-free tests and two mutants hold both.
• 230324c — the polling-select killer let the order two tasks first ran in decide whether the poller ever gave its worker back, so under the mutant one Linux run in five finished: the sender now waits for the poller, and five runs are required.
• b66f50f, 3b1a500 — thirty-four catalogue entries the contained campaign could not decide: every declared killer passed on the mutant on Linux. Sixteen saw their defect on macOS only (glibc keeps a freed block’s bytes; AArch64 does not fault on MIN % -1; one oracle sum let two wrong remainders cancel) and got killers that fail on both; nine survived on macOS too, because N90, N104 and N105 had moved what they attacked — dead code deleted (Index::choose), a rule N105 decided twice decided once, a test added for an origin no program reached, for a requirement filter and for a statement’s debug line; one entry retired as equivalent (n62-every-declaration-is-needed: since N105 a unit declares only what it calls); eight are observable only with macOS tools or Apple’s clang and are verified on the host.
• c1aef4a, edc31b5 — N104’s migration had left three concurrency programs and one scheduler test wrong; edc31b5 also, by mistake, carried the deletion of Cranelift’s old lowering, so that commit does not build alone (recorded in N105’s commit).

Mutation catalogue accounting

Every one of the catalogue’s 1,268 entries has a verdict, and every verdict is caught:

0 survived, 0 timed out, 0 runner errors, 0 not injected, 0 without a verdict.

Evidence index

target/gates/n108/evidence-82936f6/ — an evidence bundle (nazm.evidence/1), 31 files indexed by BLAKE3, index.json SHA-256 83c1463ab7c51d83a979857199f3ef13c6d579022f30d23acf02dbaa5f3e4ef6, verified with cargo xtask evidence --verify: the release’s MANIFEST.txt and PROVENANCE.txt, every gate stage’s log at 82936f6 and at 230324c, the six campaign journals, the deferred list, the host precheck and the Linux diagnoses. The artefact itself, nazm-release-candidate.tar.gz (7.4 MB), is under target/gates/n108/wt-final/target/release-candidate/, which is build output and not tracked.

Independent redigest. The archive was extracted on the host and every file’s SHA-256 recomputed with Python’s hashlib, not with the tool that wrote the manifest: 206 of 206 match, the only file not listed is MANIFEST.txt itself, and bin/nazm’s digest is the manifest’s binary-sha.

Known limitations

limitations.md is the list. In the release’s terms: the eight class-B rows’ limitations; the class-C blockers above; the oracle runs the sequential subset only; the interpreter and the compiler written in Nazm keep a thread per task; x86_64-unknown-linux-gnu is compile-only here; the archive is not bit-reproducible (it embeds its run date) and its executables were not tested for it.

The human release procedure

Every step is a person’s, in this order, from a clean checkout of exactly this commit. Stop at the first step that does not pass.

1. Check out the commit. git checkout 82936f6d857eca7e3620793655a83b39458652a9; git status --porcelain must print nothing.
2. Build the images (network here and nowhere after): docker/contained.Dockerfile (nazm-contained:1.98.1), and for the platform runs nazm-qemu:n86, nazm-wasm:n64, nazm-evm:n70 and nazm-debug:n58.
3. Run the gates, contained and offline, each exiting 0: cargo xtask contained tests; cargo xtask contained selfhost; cargo xtask contained bootstrap; cargo xtask contained bench; cargo xtask contained fuzz 300; cargo test -p xtask --test lifecycle -- --ignored --test-threads 1; cargo xtask check and cargo test -p nazm-cli --test docs; and the campaign, cargo xtask contained --profile p4t4 mutate --strategy targeted --verify-killers, resumed with --resume until complete — 0 survived and no entry without a verdict, the eight host-only entries checked on macOS.
4. Run the platform tests on the host: cargo test -p nazm-cli --test accel --test boards --test contract_property --test debugger --test debugger_v3 --test evm --test freestanding --test realtime --test wasm --test wasm_contract -- --ignored.
5. Assemble the candidate: cargo xtask contained --profile p4t4 release (one CPU does not fit the runner’s 2,400 s since the suite grew; four did, in 1,716 s). It tags, pushes and uploads nothing.
6. Verify independently: extract the archive and compare every file’s SHA-256 with MANIFEST.txt, both ways; then cargo xtask evidence --verify over the gate’s evidence bundle.
7. Decide. Only then, and only by a person: tag, push and publish, with this file beside the artefact. None of this has been done.

Appendix — the workspace suite by crate (added in R1, 2026-10-07)

Appended, not edited: nothing above this heading changed. crates/nazm-cli/tests/counts.rs requires every release record to state its suite total and a by-group table that sums to it, and this record, written after the gate, had only the one-line total in the table above — so the docs suite failed on main from the records commit on, which the gate, run on 82936f6 itself, could not see. These are the same run’s numbers, read from its log (tests.log in the evidence bundle, evidence-82936f6/gate-final/), binary by binary, grouped by crate; doc-tests are one row.

2476 passed, 0 failed, 47 ignored.