[cargo-affected-bot]

Regenerates the nine tend-* workflows with tend 0.2.13, up from 0.2.11. The diff is the action pin on every workflow plus the uv version tend-mention and tend-notifications install (0.12.13 → 0.12.17); no job, trigger, permission, or concurrency change.

What the upgrade changes for this repo:

• A cancelled gating check stops reading as green. poll_pr_checks.py derived "green" by subtraction, so a check that completed without a usable conclusion — the shape a concurrency cancellation produces when a second push supersedes the first — landed in neither the pending nor the failed bucket and the poll exited 0. Green is now an allowlist (SUCCESS, NEUTRAL, SKIPPED), and CANCELLED, STALE, or any conclusion GitHub adds later exits 2, which monitor-ci already reads as unverified (fix(poll): report a cancelled gating check as unverified, not green tend#1323). The reviewer-facing half moved with it: the approval path now names such a check as unverified beside the verdict instead of assuming a replacement run covered it, so an approval no longer reports the commit green when nothing actually passed. That matters here because every bot PR on this repo is polled against a pinned commit.
• /tmp is writable inside the sandbox. It was read-only, so a scratch path a session composed by hand under /tmp failed with Read-only file system while $TMPDIR and the checkout worked. The runner-owned containers moved to /var/tmp and stay unwritable from inside (Make /tmp writable inside the sandbox tend#1324).
• A transient sandbox-launch failure no longer burns a whole run. The SRT bridge-socket wait has a 600 ms budget that a loaded runner can miss; the launch is now retried up to three times instead of ending the run before the agent starts — zero turns, nothing posted (fix(sandbox): retry the SRT launch past a transient bridge-socket failure tend#1327).
• CI sessions no longer inherit the bot account's claude.ai skills, plugins, and MCP connectors. Claude Code syncs those into signed-in sessions; the action now writes syncClaudeAiSkills: false, syncClaudeAiPlugins: false, and disableClaudeAiConnectors: true, keeping the instruction and tool sources to the composite action and the tend marketplace (fix(claude): keep the bot's claude.ai account out of CI sessions tend#1328).
• Four agent-instruction corrections. run-tend now names .agents/ among the instruction paths restored from the base branch on a PR, and states the restore as a condition on the checkout rather than an event list that had gone stale for relayed review events — both errors told a session a path was unpinned when it was pinned, which is the direction that loses a contributor's edit (fix(skills): correct run-tend's list of restored instruction paths and events tend#1311). ground-claims now says to read a past run's code at that run's head_sha rather than at the current default branch (skills(ground-claims): read a past run's code at that run's commit tend#1319). review-runs gained the case where a run parked in waiting holds a concurrency group forever and the workflow silently stops running (skills(review-runs): find a workflow wedged behind a run parked in waiting tend#1320). post-to-github gained the matching carve-out to the #L-link rule for a body that explains a past run (skills(ground-claims): read a past run's code at that run's commit tend#1319).
• The claude CLI pin moves 2.1.270 → 2.1.278. A default on the composite action rather than anything in these workflows, so it is invisible in this diff, but it is the agent runtime every tend session here runs on (chore: bump claude-code to 2.1.278 tend#1330).

Upstream comparison: max-sixty/tend@0.2.11...0.2.13

The copy-on-write sandbox view shipped in 0.2.12 and was reverted before 0.2.13 after it broke git add -A in the agent's checkout, so it is not part of this upgrade (max-sixty/tend#1339).