fix(curator): bump skill usage when cron jobs load skills

[cdanis]

Problem

The curator archives agent-created skills based on activity timestamps in .usage.json (last_used_at, last_viewed_at, last_patched_at). These counters are only bumped during interactive sessions when skill_view() or skill_manage() is called.

Cron jobs that reference skills via their skills: config never bump these counters. The skills loader in cron runs does inject the skill text, but it doesn't call the telemetry hooks (bump_use / bump_view). Result: a skill actively used by a cron job can be auto-archived by the curator as "stale".

Reproduction

1. Create a skill with scripts referenced by a cron job
2. Configure the cron job with skills: ["skill-name"]
3. Never load the skill in an interactive session (so usage counters stay at 0)
4. Run the curator — it will archive the skill despite active cron usage

Example

The `gloom-index-calculator` skill was archived by the first curator run on May 1, 2026 even though it had an active cron job using its scripts. It was one of 36 skills pruned for "staleness" during that pass.

Root Cause

In `agent/skill_loader.py` (or wherever the skill text injection happens), when a cron job loads attached skills via the `skills:` config, the code reads SKILL.md and injects it into the prompt but doesn't call `tools.skill_usage.bump_use(skill_name)`.

Fix

Call `bump_use(skill_name)` in the skill loader path for cron jobs (and any other non-interactive context like background tasks or subagents). This ensures the curator sees these skills as actively used.

[cdanis]

Another thing worth noting: the LLM consolidation pass has no minimum age protection for skills. Unlike auto-transitions (which use created_at as a timestamp anchor so new skills do not immediately archive), the forked curator agent gets a raw candidate list and is told to consolidate/prune based on content overlap and usage counters — with no check on skill creation date.

A freshly created skill that is only used via cron (and has zero usage records due to the bump bug) can be pruned within days of creation with an empty reason field (source: "no-evidence fallback"). This happened to ymca-registration-portal-scraping — created April 25, pruned May 1.

Fixing cron bump_use addresses the symptom but does not solve this structural gap. The curator prompt or candidate list builder could add a minimum-age guardrail (e.g., "DO NOT prune skills younger than 7 days unless clearly duplicated").