[justinjoy]

Scope (PR CI only)

• .github/workflows/ci-pr.yml: all ubuntu legs → ubuntu-26.04 / ubuntu-26.04-arm (runs-on, job names, matrix, comments); 7 Set up Meson 1.12.0 step blocks removed — every meson-invoking job now installs the distro-provided meson per OS (apt ×7, brew ×2, choco ×1).
• .github/workflows/lint-pr.yml: runs-on → ubuntu-26.04 (job names untouched — branch-protection contexts).
• Lockstep updates (required, verified): 6 ci-telemetry-fixtures/run-*-jobs.json + test-ci-run-telemetry.py (test reads the live ci-pr.yml), test-ci-pr-build-timeouts.py, test-early-size-gate.sh, RELEASE_PROCESS.md required-context list, collect-size-attribution.py policy-identity token (+ clarifying comment), 2 comment-only arm renames.
• Intentionally untouched: setup-meson action (still used by android/ios/size-attribution/ci-main), ci-main.yml, release-tag.yml, tier1-sanitizers.yml, perf/fuzz/size-attribution workflows, frozen size-baseline records (verify-size-baseline.py, test-size-baseline-provenance.py).

Live-run fixes (first PR CI run 40888053298)

The first live run validated the 26.04 legs (all started and passed: TSan, TSan-native, Sanitizers/macos, both lints, RC gate — runner-label risk fully retired) and exposed exactly two distro-meson incompatibilities, both fixed in 9a62e4ea:

1. meson test --exclude does not exist in distro meson 1.10.1 (unrecognized arguments: --exclude killed the sanitizer-matrix leg). The step now derives "everything except sbom_snapshot" at run time from meson introspect --tests builddir-san (positional dir in 1.10.x) + grep -Fxv sbom_snapshot, so the other sbom-suite contract tests still run (which is why --no-suite sbom was the wrong tool). test-required-gates.sh gains a sanitizer_excludes_snapshot_dynamic assertion for the PR workflow; ci-main.yml and tier1-sanitizers.yml keep the original --exclude assertion untouched (they stay on pinned meson 1.12 via setup-meson).
2. SBOM baseline drift — removing the setup-meson action references from ci-pr.yml dropped the scanned local-action entries (12 → 11 workflow references); sbom/snapshot.txt regenerated with CI's pinned syft 1.50.0 and now byte-matches a clean CI checkout scan (verified by running the gate's exact pipeline; the failed run's own diagnostic showed 11).

Live-run fixes (second PR CI run 40932685444 — 18/19 green)

Every ubuntu-26.04 leg, both macOS legs, and all lints/gates were green; the only failure was Build / windows-latest / msvc: choco install meson succeeds (official 1.12.1 MSI) but the MSI writes its PATH entry into the machine registry, which the job's already-snapshotted environment never sees, and chocolatey creates no shim for MSI packages — so meson was unresolvable at Configure ("The term 'meson' is not recognized", job log 40932685444). Fixed in 68dcd011 (.github/workflows/ci-pr.yml, Windows install step only):

1. Resolve the MSI install dir explicitly — chocolatey's own HKLM:\SOFTWARE\chocolatey\installed\meson record first, then the MSI's documented default locations (C:\Program Files{, (x86)}\Meson{, meson}), then a bounded scan as last resort; throws if nothing is found (fail-loud).
2. Verify before publishing — meson.exe --version and ninja --version executed by absolute path (or shim) with $LASTEXITCODE checks, because GITHUB_PATH only takes effect in later steps; the verified dir is then appended to GITHUB_PATH — the same mechanism the former setup-meson action used.
3. ninja is now explicitly in the choco install list and its resolvability is asserted, so meson setup can find it in every delivery mode (MSI-bundled beside meson, or chocolatey shim on base PATH).

Rerun note: the first attempt of this commit failed only on the pkgconfiglite download (SourceForge mirror timeout — exited 404, transient infra); the new $LASTEXITCODE check caught it and failed the step loudly, exactly as designed (meson/ninja had already landed). The targeted re-run (gh run rerun --failed) passed: 19/19 jobs green, with meson 1.12.1 resolved from the documented install dir and ninja 1.13.2 (base-image shim) verified in-step before GITHUB_PATH publication.

First run of 68dcd011 (40978095312): all 18 non-Windows legs green on the first attempt; the Windows leg failed only because the SourceForge mirror for pkgconfiglite timed out mid-download (transient infra, exited 404) — the new $LASTEXITCODE check caught it and failed the step loudly, exactly as designed. The targeted re-run (gh run rerun --failed) then passed: 19/19 jobs green, with meson 1.12.1 resolved via the documented candidate dir and ninja (base-image 1.13.2 shim) verified in-step.

Accepted trade-offs (documented)

1. Meson version drift — PR legs now run the distro meson (unpinned), as directed. Mitigants: meson.build only requires >=0.62.0; gate scripts are version-tolerant by design (check-shell-gate-walltime joins on command, check-required-correctness accepts both name forms); the two "Verified on meson 1.12.0" comments are reworded to historical-verification phrasing.
2. ninja parallelism cap (setup-meson/ninja-cap.py, NINJA/WIRELOG_NINJA_REAL) dropped from PR legs — zero consumers in ci-pr.yml/meson.build; accepted.
3. Runner labels — ubuntu-26.04 / ubuntu-26.04-arm: confirmed present by the first live run (every 26.04 leg scheduled and started).
4. Follow-up hardening (accepted residual, non-blocking): the dynamic-exclusion gate assertion uses substring matching (a superstring exclusion target would still pass; worst case is the redundant SBOM scan running under sanitizers → visible timeout, never a silent skip), and the introspect pipeline has no pipefail (broken introspect ⇒ empty list ⇒ full suite runs, fail-visible). Both bounded; candidate for a later PR.

⚠️ Maintainer action required

Re-point branch-protection required-check contexts to the new job names, e.g. Build / ubuntu-latest / gcc → Build / ubuntu-26.04 / gcc, Build / ubuntu-24.04-arm / gcc → Build / ubuntu-26.04-arm / gcc (full list already updated in docs/RELEASE_PROCESS.md). Until then, old-name required checks will block merges.

Size gate note

PR size gate is advisory mode; the baseline (428046) was calibrated on the 24.04/gcc-13.3/meson-1.12 profile — re-verify on the 26.04 profile before flipping to enforced.

Validation (all green, uv venv)

test-required-gates.sh (26 assertions), test-gate-skip-exit-codes.sh, test-early-size-gate.sh, test-generate-sbom.sh, test-sbom-venv-exclusion.sh, check-sbom-wrap-pin.sh, check-sbom-snapshot-locale.sh, check-sbom-snapshot.sh (118 packages, syft 1.50.0), test-check-shell-gate-walltime (17), test-ci-run-telemetry (41 + subtests), test-size-attribution (41), test-branch-protection-contract (13), test-tsan-timeout-policy (11), test-size-baseline-provenance; all three workflow YAMLs parse. The meson-1.10.1 mechanics were additionally exercised end-to-end on a scratch project by the review roles (--exclude rejected as in CI; positional introspect + exact-name meson test selection works; other sbom-suite tests retained). Pre-existing base failures (not caused by this change, verified on clean tree): test-ci-build-cap 1 fail (path canonicalization), test-required-correctness 1 error (subprocess launch).