fix(skill): retry registry fetches and tolerate blocked backup cleanup

[gujieye]

Summary

Fixes two issues in bl skill init:

1. Index fetch has a 10s timeout and no retry
   • INDEX_TIMEOUT_MS raised from 10s to 30s; fetchSkillsIndex / downloadSkillAsset now go through withRetry (3 attempts by default; network/timeout/5xx/404 errors are retried)
   • Added an optional attempts parameter so the silent advisor background channel passes 1 and stays fail-fast, without stalling bl advisor recommend
2. Upgrading a large skill is misreported as a failed install by host safe-delete guards
   • The old bailian-docs-llm-wiki has 1583 files. After atomicSwap completes, deleting the .old-* backup trips the host guard (">500 files deleted per turn requires confirmation"), and the error bubbles up: an install that already succeeded is reported as failed and the lock entry is never updated
   • Backup deletion after a successful swap is now best-effort (try/catch): when the guard blocks it, the install still succeeds. Leftover .old-* dirs are inert (status scans already ignore them). The duplicated logic in postinstall.js is fixed the same way
   • Also made temp-dir cleanup best-effort so cleanup failures no longer mask the real error

Test plan

• 5 new cases in skills-installer.test.ts: guard-blocked backup deletion still installs, cleanup failure does not mask the original error, index fetch retries then succeeds, attempts=1 disables retries, asset download retries transient HTTP errors; existing 404 case now asserts the 3 retries are exhausted
• vp check passes (0 errors)
• skill e2e 12/12 passed (including add/list/remove lifecycle against the real registry)

Known issues

advisor-sync (1 case) and skills-agents (8 cases) fail on this machine — pre-existing environment issues (/etc/codex and other real agent dirs on this host break detection expectations), reproduced on unmodified code, unrelated to this change.