docs(handoff): skill-usage-telemetry closed out — both PRs deployed, plus a measured store-api flake and a worktree-upstream trap

[ZacxDev]

Final handoff update for the skill-usage-telemetry effort.

Closed out: #1119 (the RESULT: verdict-grammar guard, 17 audit rounds, closed clean) and #1150 (the close-out) are both merged, verified on origin/main by content, and deployed — ship.sh run twice, both hosts converged and cross-host COMPARED at d9144e21.

🔴 One gap this fixes, which I had left myself: I merged #1150 and did not ship, leaving the laptop 6 commits behind. That is the merged-≠-deployed failure this repo documents; now closed.

New open investigation, with the measured evidence: scripts/tests/test_subsystem_store_api.py is a fleet-wide flaky gate — red on 4 of 12 open PRs in one sweep, each naming a different test in that file, 7 others green. The discriminating control is recorded because it settles attribution in one run and needs no second PR in flight: the same sandbox tier CI uses passes locally on the branch (collected=19913 failed=0), including the exact test CI failed on. Same code + same tier + different result ⇒ environment, not diff.

This is a recurrence of an already-recorded open item (the 2026-08-28 index bullet, CLOSES WHEN: green 5/5 under induced load), so it is ranked as forcing: gate rather than filed as a new discovery.

Gotcha worth having in writing: git worktree add -b <topic> origin/main sets the new branch's upstream to origin/main, so a bare git push from that worktree targets main. Mine silently did nothing (branch protection), and git status -sb read ## <topic>...origin/main [ahead 2] — which looks like an ordinary ahead-count. Caught only by comparing rev-parse HEAD against rev-parse @{u}.

The activity_reader credential rotation remains the only item with an external forcing function, and it is Zach's.

🤖 Generated with Claude Code

https://claude.ai/code/session_01J2CefG8pW2c1hRA9agpsUr