Quirl 0.3.0 native release evidence

Recorded 2026-09-06 for the published v0.3.0 release. All four native artifacts passed the enforcing release benchmark with 101 successful PTY samples each. These results measure the exact published executables on the hosted runners below; they do not establish universal latency guarantees.

Release identity

• Candidate and resolved tag commit: dfb43e54320d166ebe195bb5a0255b8aacbb11e5.
• Native build and performance run: 34054071682, attempt 1.
• GitHub release ID: 383694213; published at 2026-09-06T19:34:54Z as a nondraft, non-prerelease, immutable release, verified as latest at publication.
• Release manifest SHA-256: 583a36d0d4ae5ae85e336d71695480bee2624778a41f7f52ff7bdd3ec2820329.
• All 15 published asset digests and byte counts, and the exact release body, matched the independently verified release bundle. Each archive executable matched its native benchmark report.

The binary and benchmark independently reported this same clean source revision. Artifact digest, profile, source identity, and harness identity checks passed on every target.

Native results

All values below are milliseconds except executable bytes. Limits were startup P50 ≤25 ms, first-prompt paint P95 ≤21 ms, and keystroke-to-frame P95 ≤8 ms. Each row completed 101 of 101 requested PTY samples with no recorded sample failure.

Size follows ADR 0036: the project has no default executable-size ceiling. All four sizes exceeded the advisory 8 MiB threshold and retained warnings. Schema 8 records hard_ceiling_bytes, enforced_limit_bytes, and hard_gate_passed as null; exact measured bytes remain present. No caller maximum was requested. Latency, identity, sample completeness, stream retention, and cleanup remained enforced.

For actual timing measurements, schema 8's per-metric release_gate_accepted records valid measurement evidence. Numerical acceptance is separately represented by target_result and the overall performance_gate_passed; a valid measurement that misses a budget still fails the overall gate. All four published reports pass both validity and numerical checks.

Runners and toolchain

The Linux ARM64 CPU model was unavailable to the harness and remains unknown. Exact hostnames, memory counts, timestamps, and toolchain output remain in the original JSON reports. macOS ARM64 used a virtual Apple M1.

All builds used Rust 1.97.1 (8bab26f4f, 2026-07-14), LLVM 22.1.6, and Cargo 1.97.1 (c980f4866, 2026-06-30). The official release profile used optimization z, fat LTO, one codegen unit, stripped symbols, and panic = "unwind".

Exact report files

These are byte-for-byte copies of the schema 8 reports uploaded by the native jobs. Their SHA-256 values identify JSON bytes, including whitespace, separately from executable and archive hashes.

Published executable hashes

Each digest below was checked against the actual bin/quirl bytes in its tar archive. Each published archive digest and byte count then matched the release manifest and GitHub asset metadata.

Preserved earlier attempts

Earlier candidates did not publish a tag or release. Their failures remain separate from the published candidate:

• Run 34033659127, candidate d55033b703314e25d6ec69c9faedf55352f510af, and run 34035129106, candidate c45ae469b057056c949ba21d502421c04d7cc075, failed the initial canonical gate before packaging.
• Run 34036368955, candidate bee307192d55c7d69058410b09f5be76b322c9bc, passed three native targets but failed Intel first-prompt P95 twice: 28.309796 ms on attempt 1, then 22.811006 ms on its single bounded attempt 2, against the unchanged 21 ms limit. Both completed all 101 PTY samples. The executable, archive, provenance, methodology, and thresholds matched byte for byte across those attempts. Both original failure reports remain unchanged.
• Run 34040112452, candidate d2d5070e271adb5a1ada5ca4ac987cb63d2adad9, failed a dialect-island cancellation PTY check before packaging. Its captured screen showed the 30-second execution deadline and status 1 instead of the expected cancellation state. No native performance artifacts were produced.

The two failed Intel reports have SHA-256 values cb2bcc55eeff8dbd1b1a5c02742168ee071b0cfd5d5924a8af61db8a364f59e5 and 7c2588d66bf52c2d80f702f033d1d387a793ac15eec69d3ddbd8e101f0651111, respectively. Their timings demonstrate variability for identical bytes, not an infrastructure-only cause. The published candidate's passing measurements do not retroactively pass earlier candidates or isolate the cause of every difference. No size or latency budget was relaxed to obtain this result.

Distribution and recording

The composed release run completed publication and its protected Homebrew job successfully. Tap PR 8 merged as 57a937fe2b674dd29e7475522889f3bc77c88711; the formula on tap main matched the reviewed four archive URLs and hashes. The website asset run 34055349465 passed from the released candidate and produced asset evidence commit cffebda4b43ad57f18dd3c1e63260ef381c059af.

The fresh terminal recording and provenance use the exact published macOS ARM64 executable, SHA-256 30a6edce509a6220cfb4c505fe0192598d56174336a8720ffd0e2042be1c53d8. VHS recorded 71.12 seconds in an isolated real PTY with key input; the source revision, package, report, runtime assets, media digests, and sampled visual review are recorded separately. These local recording timings are not benchmark measurements. Provenance explicitly records no human or physical-terminal QA signoff.

Method and limitations

The PTY harness uses a private initially empty home, configuration, history, catalog, project database, and XDG directories, shared across its samples. Each sample starts a fresh native process in a 120×40 terminal, answers cursor-position queries, reconstructs the screen, checks the prompt and exact binary identity, then proves editability. Startup ends at the editable-marker frame; first-prompt paint and the final representative keystroke use separate endpoints. Percentiles use nearest-rank selection.

After the measurement endpoints, successful samples clear the input and request normal EOF exit within the original two-second sample deadline. Normal exit and bounded process cleanup finish before the next sample. Shutdown and screen stabilization are excluded from first-paint latency. Failed samples retain forced cleanup and cannot produce a passing partial report. Native job logs retain all 101 ordered timing and shutdown observations per target.

Hosted scheduling, virtualization, CPU frequency, other machine load, and filesystem caches are not controlled. The reconstructed PTY frame does not measure physical terminal-emulator scheduling, GPU composition, or monitor scanout. This automated evidence does not cover every terminal, visual accessibility scenario, clipboard, IME, or sustained human session. Separate contract, cancellation, restoration, and long-session checks complement it; no human checklist item is inferred from these results.

Historical 0.2.0 measurements and earlier release research retain their original artifact identities, methodology, limits, and outcomes.