chore(release): 0.62.0 — gate integrity

[noahgift]

v0.61.0 fixed things that were silently broken or silently false. This release goes after the machinery that was supposed to catch them.

Almost every entry is a gate that could not fail, or a number that was not true — and the pattern is worth naming: every one of these was green beforehand.

• A substring pass-grep matched its own failure text (0 error matches 10 error(s))
• A pipeline reported tee's exit code instead of the test's
• A parity gate compared one token
• A skipped result set passed: true
• A coverage job measured a facade and reported 0%
• 547 integration targets were never executed by any workflow; 11 no longer compiled
• The nightly regression story died six lines in, while its own alerting path had never once worked

Three speed claims re-measured; none survived

claim	reality
GaussianNB "4.9×"	a dev-box number; re-pinned to the host that runs it (#2328)
LinearRegression ratio	was measuring the host, not the algorithms (#2345, #2346)
ollama GPU decode "1.371×"	withdrawn — measured reality on sm_89 is 1.015–1.109×, i.e. parity (#2348)

The last is withdrawn rather than re-explained: it was measured under a Q4K kernel default that no longer exists and cannot be reproduced on the host. Under-claiming is also a reporting failure, so the replacements carry host, spread and remaining headroom instead of being rounded up.

Verification before cutting

cargo set-version --workspace 0.62.0   exit 0, 35 manifests + Cargo.lock, no half-bump
all 78 crates                          at 0.62.0, none left behind
cargo fmt --all -- --check             exit 0
cargo test -p aprender-contracts       1420 passed / 0 failed
cargo audit -n                         exit 0
cargo deny check advisories            exit 0, "advisories ok"
cargo metadata --no-deps               exit 0
scripts/check_readme_claims.sh         4/4 PASS (78 crates, 1767 contracts, 103 commands)

The v0.61.0 set-version blocker at Cargo.toml:509 stays fixed — no non-exact requirement, so no half-bumped tree.

Known, not blocking

Two nightlies went red in a 15:20–15:40 window caused by restarting 16 previously-stopped CI runners, which spiked mac-server to load ~27:

• GaussianNB beat 0.535 > 0.50 — the harness's own discriminator says contention (fit and predict both up; apr 64.8 ms vs its 27.9 ms idle reference). fix(beat): the GaussianNB "4.9x" was a dev-box number - re-pin to the host that actually runs it #2328 deliberately declined to weaken this gate.
• Coverage Nightly reported an empty percentage — a shared-CARGO_HOME race on cargo-llvm-cov, filed as Coverage Nightly races on the shared ~/.cargo/bin (cargo-llvm-cov 'No such file or directory') — security job already has per-job CARGO_HOME, coverage does not #2353 with the fix (the security job already uses a per-job CARGO_HOME).

Neither is a code defect. Both will be re-run to confirm green before the tag is pushed.

🤖 Generated with Claude Code