fix(skills): persist quarantined trust row for AutoSkill drafts at generation time

[bug-ops]

Summary

SkillGenerator::write_quarantined wrote an AutoSkill trace-extraction draft's SKILL.md to _quarantine/<name>/ but never persisted a skill_trust DB row. The doc comment claimed the _quarantine directory prefix kept the registry from ever loading these drafts, but the loader applies no directory-name filtering at all — quarantine was never actually enforced at the trust layer. Because no DB row existed yet, the draft's first hot-reload had no existing row to preserve and fell through to [skills.trust] default_level, silently trusting an unreviewed draft whenever an operator set default_level = "trusted". Separately, AutoSkill-generated names had no collision check against native tool IDs.

• trace_extraction::log_and_persist now eagerly writes a Quarantined skill_trust row right after generation, keyed on the name actually written to disk (both Add and Merge branches), before the skill can ever reach a hot-reload.
• A guard skips the write (and warns) instead of overwriting when the target name already has a trust row pointing at a different on-disk path — the merge LLM is prompted to preserve the existing skill's name, so a merge landing on an already-Bundled or -Trusted skill is the dominant case, not an edge case, and must not clobber that skill's trust row.
• The native-tool-ID name-collision warning is extracted into a shared helper called from both the startup catalog emit and every hot-reload, so it no longer depends on SemanticMemory being configured and fires at startup rather than only after the first post-startup file change.
• Corrected the false _quarantine "security boundary" doc comment on write_quarantined.

Closes #6702

Test plan

• cargo nextest -p zeph-skills: 590/590 passed
• cargo nextest -p zeph-core: 2232/2232 passed (2 skipped)
• cargo nextest --workspace --lib --bins (CI feature set): 15269/15269 passed
• cargo +nightly fmt --check, cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings, rustdoc gate, gitleaks protect --staged all clean
• New regression-lock tests: eager quarantine survives a hot-reload despite default_level = trusted; merge into an already-Trusted skill resets to Quarantined; merge/add against a differently-pathed Bundled/Trusted skill leaves its trust row untouched and warns instead; native-tool-ID collision warning fires with and without SemanticMemory configured, from both the startup and hot-reload call sites
• .local/testing/playbooks/autoskill-a1-a2.md updated with 3 new live-session scenarios (not yet run live this cycle) and a corrected "Known Edge Cases" note; .local/testing/coverage-status.md row updated in place