Staleness behaviour is calibrated per platform

Summary

The remainder of itd-111 that spc-22 did not deliver. spc-22 closed on 2026-09-23 with acceptance criteria 1 to 7 delivered: the SessionStart staleness notice, the install refusal on a stale or unknown vintage, install mode and vintage on version and ahoy, no version-discovery network request without version --check, the transition report, and the unknown-never-fresh outcome. This spec carries what did not ship.

The delivered part was already announced in the [0.5.0] changelog section, and the intent carries no shipped_in: because it stays planned. The close that ships the intent through this spec is the one that decides how the cut reports it: without shipped_in: the whole intent is announced again, which is the redundant line the changelog composer prefers to a silent omission.

Scope

• Acceptance criterion 8 (missing). Given the same staleness scenarios on macOS and on Linux, when the itd-109 calibration runs, then observed behaviour matches, recorded as a machine-class criterion, human-verified per platform. No recorded macOS or Linux human-verified calibration run exists on main at closure.

How the criterion is satisfied

The product thinker ruled on 2026-09-29 that an agent's check on macOS plus the Linux CI run stand in for the per-platform check by a person, and that the intent may be finished. The criterion is amended to match in the intent, which states the ruling. itd-109 is still a draft, so the calibration harness the original wording named does not exist; the scenarios run through the tests and binaries that already carry them.

Linux: CI. The push run on main at de42f275a (the merge of pull request #751), CI workflow run 36608836118, job check (ubuntu-latest) (job 109544802516), built with go1.26.7 linux/amd64 and passed internal/core/vintage, internal/core/ahoy and internal/surface/cli under go test ./... and again under the race lane. None of the staleness tests below skips on Linux.

macOS: the agent's check, 2026-09-29, at the same commit (darwin/arm64).

1. The 25 staleness tests below passed under the declared toolchain (go1.26.7) and under the local go1.27.1, zero skipped, in 4.7 seconds wall time with -p 4: TestOfReaderReadsABinaryWithoutRunningIt, TestCheckoutTipProvider, TestPinnedVersionProvider, TestGitHubReleaseFetcherReadsTagFromRedirect, TestGitHubReleaseFetcherNoTagIsAnError, TestReleaseProviderComparison, TestCompareOutcomes, TestCompareReportCarriesValues, TestCurrentFromSettings (vintage); TestVintageFromDogfoodCheckout, TestVintageFromUnknownCurrentIsTerminal, TestVintageFromPinnedInstall, TestVintageDisplayAndStaleness, TestReadPinnedTagPrefersRootThenCache, TestDogfoodStalenessFrom, TestInstallRefusesUnknownVintage, TestInstallRefusesStaleAgainstTip, TestInstallOverrideProceedsThroughStaleBinary, TestInstallProceedsThroughFreshBinary (ahoy); TestOnlyUpdateCheckTouchesTheNetwork, TestUpdateCheckNamesTheNextStep, TestSessionStartReportsVersionTransition, TestSessionStartSilentWhenVersionMatches, TestFormatStalenessNotice, TestModeAndStatuslineTouchNoNetwork (cli).
2. End to end, on a scratch copy of that tree committed into its own repository, with a VCS-stamped binary built from it and a throwaway home folder:
   • Fresh. abcd --version and abcd ahoy print the binary's revision and up to date; abcd hook session-start gives no staleness notice.
   • Stale (one commit added after the build). Both print stale — behind the checkout tip naming the tip; the session-start notice names the binary, the commit it was built from, the tip, and make build; abcd ahoy install refuses with the mismatch named, and neither the home folder nor the working tree changed.
   • Unknown (a dirty rebuild, vcs.modified=true, and a -buildvcs=false build). Both print unknown, never up to date; the session-start notice names the dirty rebuild; abcd ahoy install refuses naming the rebuild fix, writing nothing.
   • Explicit check. abcd update --check fetched the latest release once and reported it with its source named (the check abcd version --check became when the verbs consolidated, itd-2609212130136102). Each surface answered within 0.4 seconds; the session-start hook, which also drains transcripts, within 0.2 seconds.

The observed behaviour matches across the two platforms. The behaviour itself shipped in v0.5.0 and was announced there, so the intent carries shipped_in: v0.5.0: this close adds verification and no new behaviour, and an announcement at the next cut would repeat the 0.5.0 line.