[screenleon]

Summary

Fixes two product defects that only show on native Windows, makes three suites pass there, and adds CI that runs suites natively.

• Memory directory not found. Claude Code on Windows names a project directory after the drive-letter path with every non-alphanumeric character turned into - (C--Users-Jo-Doe-repo). find_memory_dir (hook lookup, memory.sh) and the installer/migrator lookup (memory-dir.sh) only tried the POSIX spelling, so a repository on a native Windows host never found its memory. Both now also try the Windows name; other platforms are unchanged.
• Memory ranking. A native sqlite3 ends rows with CR LF, which left a CR on last_access_day, read as 0, so every card ranked oldest. The reader strips it on native Windows only.
• Tests. test-pmctl-memory, test-pmctl-pm and test-pmctl-ship-finish now pass natively (default /tmp and a clean scratch dir): fixtures use the spelling the product records, the scratch root is the physical path, and the ship-finish bash -c shells load the jq LF shim the product entry loads. Cases that need tab/backslash file names or real symlinks skip with a stated reason.
• CI. New windows-native-suites job (windows-latest, one job per suite, 8 suites) beside windows-native-smoke.
• Gate quota detection. A real gate that hit the Codex usage limit reported an unexplained "transport failure" and retried: a run through pmctl dispatch run has no adapter banner in the reviewer log, only run: <id> and a <id>.footer naming the executor's stderr/event files. The check now follows the footer (and still reads a banner in a directly dispatched log).
• Docs (docs/platform-support.md) and CHANGELOG updated.

Verification

• The last PR gate attempt on an earlier commit ended without a verdict because the Codex plan quota was used up (resets 2026-11-12 11:20); the reviewer is not switched. A GO on the final commit is still required before merge.

• PR gate on the earlier commit eee228e, native Windows, default environment (no PM_DISPATCH_STATE_ROOT, no TMPDIR): GO; verify artifact_valid/subject_current/policy_applicable all pass. Earlier rounds on this branch found a real fixture defect under the default /tmp (fixed) and the sibling lookup in memory-dir.sh (fixed).

• WSL2: test-pmctl-memory 138, test-pmctl-pm 32, test-pmctl-ship-finish 82, test-guards 329, test-doctor 97, test-pmctl-context 206 all pass; CI lints 19 passed, 0 failed, 2 skipped (the skipped ones need CI-only variables).

• Not verified: a scratch directory on a non-C drive (this machine has only C:), and the new windows-native-suites job on a real runner until this PR's CI runs.

• Native Windows stays experimental; promoting it needs this job to stay green first.

• Permanent test admissions: one case added to the pre-existing test-pmctl-memory.sh (memory discovery: both helpers find the native Windows project directory name): criteria 1-6 met (it pins the observable result of both public lookups for the documented Windows project-directory name, with the physical path a hook passes; no source grep, no private helper); layer: library, observing the printed directory; it applies to native Windows only and skips elsewhere with a stated reason. The gate round that flagged its fixture (hard-coded drive letter, /tmp alias) was resolved by deriving both from the real path, not by adding cases. A second case was added to the pre-existing test-gate-protocol.sh (quota signature: found through the run footer next to a pmctl dispatch log): criteria 1-6 met (it pins the observable output of the public function for the layout measured on a real failed gate, including a negative control; no source grep); layer: library. The reviewer-quota test in test-pr-gate.sh was not extended.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ESW7P8BvZ5Crn6vi6EA2J7