Step 39: the TLS brick’s corrections, and a write that reaches the disk on macOS

Two independent auditor sessions read step 37 on 18 Sep 2026 and both showed, with real handshakes through the brick’s exported ABI, that the client accepts certificate chains outside their issuer’s restrictions. They also found an ALPN boundary, a hole in the fuzz driver’s count, a corpus test file that stays green when the handshake is replaced by a Timeout, a stale spec paragraph, and the standard’s vectors missing from the corpus. All conceded (the decision-log rows of 18 Sep). The same morning the lead found that both runtimes call plain fsync, which on macOS does not ask the drive to flush. This step closes all of it. The brick is not complete until it lands.

Orientation

The auditor’s readings and their scripts, which are this step’s gate and are not to be edited: audit/mo-audit-2026-09-18-tls-independent-code.md, audit/mo-audit-2026-09-18-recent-tls-step37.md, audit/evidence/2026-09-18/tls-independent-code/chain-checks.py, audit/evidence/2026-09-18/recent-tls-auditor/{chain-checks.py,fuzz-accounting-check.py,example-assertion-check.py}. The code: toolchain/src/bricks/tls.zig (isCa near line 538, checkChain and signedBy near 648 to 731, the ALPN parsers near 357 to 410 and 1405, the record layer’s handling of a plaintext record after the handshake keys are installed), Zig 0.16’s std/crypto/Certificate.zig (what its parser gives and what it skips), toolchain/bench/step37/{fuzz.py,diff.py}, examples/effects/tls-client.mo (the faulted? and heard? allowances), mo-wiki/spec/design-v0/09-stdlib.md lines 341 to 357, toolchain/src/blocking.zig:114 and toolchain/runtime/mo_rt.c:8303 (the two fsync calls).

Write scope

toolchain/ (the brick, both runtimes, their tests, bench/step37/ and a new bench/step39/, PRELUDE.md if a row’s wording changes), examples/effects/tls-client.mo and a new examples/effects/tls-vectors.mo with their .expected and .mo.ids sidecars and fixtures under examples/effects/tls/, and these spec lines only: chapter 9’s ## Tls paragraph on what the client checks and the stale fixture paragraph (line 355), chapter 9’s ## Fs sentence on what “on disk” means per platform, and the runtime paragraph of toolchain/README.md. Nothing else under mo-wiki/, nothing under audit/.

Parts

A. The chain, checked as RFC 5280 asks for what the cut supports, and refused for what it does not. For every certificate in the path: pathLenConstraint enforced; an issuer must carry keyCertSign when it has a key-usage extension; the leaf’s extended key usage, when present, must include server authentication for a server’s leaf (client authentication is out of the cut); every critical extension the brick does not process is a refusal (unsupported_certificate or bad_certificate, say which and why); name constraints on any certificate in the path are refused as unsupported rather than ignored; validity checked at the boundary second. Each is a brick test with a generated Ed25519 chain and a P-256 chain, a test under both runtimes through TlsClient.connect over a real socket, and a case in the differential generator with OpenSSL as the oracle (the oracle is never weakened to match). Untrusted stays the row’s answer. Then run the x509-limbo corpus’s TLS-server cases that fall inside the cut (Ed25519 and P-256 chains) and report accepted-but-should-reject as a number that must be zero, and rejected-but-should-accept as a list with a reason each.

B. ALPN. Either the whole legal offer is searched or the row states a limit and an offer past it is refused with a named error at offer, the same on both ends; never a silent truncation that turns an overlap into no_application_protocol. The auditor’s 65-name case passes.

C. A client never acts on a record in the clear once the server’s handshake keys are installed (after the ServerHello): such a record, alert or not, is the client’s own failure and connect is Handshake. The server’s side keeps today’s tolerance of a client’s plaintext alert before the client’s Finished, because OpenSSL’s client sends its refusal alerts that way, and the spec’s ## Tls says so in one sentence. Step 38’s abuse table gains the distinction: cells C2 sent in the clear expect Handshake in both the alert and the close_notify columns.

D. The instruments. fuzz.py counts and keeps every failed batch (its inputs and output saved under work/failed-batches/), whether or not an input fails again alone, prints both numbers, and exits nonzero on either; diff.py exits nonzero on a mismatch; the differential generator passes its sampled groups to s_server on the client path, and draws chains from part A’s constrained shapes. Then a fresh 1,000-session differential run and a fresh fuzz CPU-hour on the corrected brick, seeds stated, raw output filed.

E. The corpus says something. tls-client.mo’s tests are split: exact deterministic controls with no fault allowance (a handshake and a line each way must succeed; each refusal must be its named error) and separate seeded tests where a fault was actually injected. tls-vectors.mo: the RFC 8448 section 3 trace as a corpus test under both runtimes (what can be replayed without the RSA certificate’s authentication is replayed and the file says what is skipped and why). Chapter 9’s stale fixture paragraph is removed.

F. F_FULLFSYNC on Darwin, both runtimes: where the runtime syncs a file on macOS it calls fcntl(fd, F_FULLFSYNC) and falls back to fsync only when the file system refuses it (ENOTSUP, ENOTTY); Linux unchanged. A test that traces the call on a Darwin build (a counter the runtime test can read, not a timing). Chapter 9’s ## Fs says what “on disk” means on each platform. Measured: jobq’s pairs a second at 32 workers and creates a second on the Mac, before and after, with measure.py at 30,000 jobs.

Numbers

The chain check’s cost per handshake before and after part A (µs, both key types); handshakes a second before and after; the differential run’s counts by refusal reason; the fuzz’s inputs, batches, failed batches, crashes; the x509-limbo counts; jobq on the Mac before and after part F; the suite’s time. The date, uptime, and the top of ps at the head of every table; best of five.

Done when

Both copies of the auditor’s chain-checks.py, unedited, show the valid control ready and all four constrained chains refused, and the 65-name ALPN case ready (raw output filed under toolchain/bench/step39/); the auditor’s example-assertion-check.py mutant now goes red; its fuzz-accounting-check.py now reports the failed batch; x509-limbo’s accepted-but-should-reject is zero; the differential run is 0 mismatches of 1,000 with the constrained chains in the draw; the fuzz hour is 0 crashes and 0 failed batches; python3 toolchain/bench/step36/guard.py 2400 -- zig build test --summary all green with its summary line and exit code in the report, on the Mac, and the brick’s own tests green on Linux (the OrbStack container step 38 used); step 38’s abuse table at 64 of 64 with part C’s cells; the numbers table; the spec lines; a numbered list “Decisions the brief did not cover”. One commit per part, subject Step 39 part X, pushed after each. Every process under the guard.

What the lead will do at acceptance, written before the work

Run the suite; run both auditor scripts from a clean clone; write one new constrained chain the brief does not name (a pathLenConstraint of 1 with two intermediates below it; an intermediate whose key usage is present and empty) and expect a refusal; break one check in tls.zig by hand in a scratch worktree and confirm a corpus test and a brick test go red (the mutant rule); read the F_FULLFSYNC call in both runtimes and trace it with dtruss or the counter.

Result at the move checkpoint: incomplete

18 Sep 2026, 4:33 PM ET, recorded by Codex (GPT-6). Robert asked to wrap the worker up ASAP. A (faec6b5) and B (18d45a6) are pushed; their worker suites report 240/240 and 242/242, exit 0. C/D are preserved in bba9fda as toolchain/bench/step39/wip-parts-c-d.patch, with continuation notes in WIP.md; E/F and Linux verification have not started.

Not accepted. Limbo is 27 accepted-but-should-reject, not the required zero. The worker’s rationale is in its report for the next lead to assess; this checkpoint changes no gate. C’s abuse table is 63/64 and its full suite is unrun. The lead’s complete acceptance list remains owed. Darwin full-sync and its Mac-specific trace/measurements are still outstanding. Raw preservation bundle: audit/evidence/2026-09-18/step-39-move-checkpoint/.