fix(skills): preserve command cache on scan setup failure

[LeonSGP43]

Summary

• build skill slash commands in a local map before replacing the global cache
• keep the previous _skill_commands mapping available when scan setup or directory traversal fails
• add a regression test for preserving cached commands when scan setup raises

Fixes #18659.

Verification

• scripts/run_tests.sh tests/agent/test_skill_commands.py::TestScanSkillCommands::test_preserves_cached_commands_when_scan_setup_fails tests/agent/test_skill_commands.py
• git diff --check origin/main...HEAD

Overlap check

• Checked open PR fix(skills): log warnings instead of silently swallowing exceptions in skill_commands.py #2868, which adds logging around swallowed exceptions but does not remove the unconditional cache clear. This PR keeps the scope to the cache-preservation bug called out in [Bug]: scan_skill_commands unconditionally clears _skill_commands before try block, silently loses all skills on scan failure #18659.

[wesleysimplicio]

Closed my competing PR (#22514) in favor of this one — same swap-on-success pattern for scan_skill_commands(). Per @alt-glitch's note, #18720 was already deferring to this PR. Looks ready; the only friction is the 12-day gap since the last push, so a rebase against current main might revive review attention. Happy to send one off your branch if useful.

[teknium1]

Thanks for isolating the unconditional-clear failure path. The bug is still present on current main: scan_skill_commands() clears _skill_commands before scan setup at agent/skill_commands.py:327-332, then swallows setup failures at agent/skill_commands.py:385-386.

Problems

• Current main added the platform-scoped cache marker in c73594fe4. scan_skill_commands() sets _skill_commands_platform before setup (agent/skill_commands.py:327), while get_skill_commands() trusts it to decide whether to rescan (agent/skill_commands.py:397-401). A command-map-only swap would preserve an old platform's mapping but record the requested platform after failure, suppressing the required retry.
• The supplied test covers only _skill_commands, not _skill_commands_platform.

Suggested changes

• Stage and publish the command map and platform marker together only after a successful scan; retain both prior values on outer failure.
• Add a cross-platform setup-failure regression that exercises the marker and subsequent rescan behavior.

Automated hermes-sweeper review.