Cap redirects in bin/install.js fetchRawText to prevent loop

[derek-palmer]

This was generated by AI during triage.

Problem Statement

fetchRawText(url) in bin/install.js follows HTTP 301/302 redirects by recursing into get(res.headers.location) with no depth limit. A redirect cycle (301 → 301 → …) or a long redirect chain causes unbounded recursion and request fan-out before any failure surfaces. Pre-existing (predates #65; the loader work only renamed get(url)→fetchRawText and moved it out of fetchSkill).

Real-world risk is low: the only caller URLs target RAW_BASE (raw.githubusercontent.com, trusted), reached only when no local checkout/packaged tasks.json or SKILL.md exists. But the function is now shared by fetchSkill and loadTaskSkillSlugs, so hardening it protects both.

Solution

Add a max-redirect cap. On exceeding it, resolve null (same failure contract the function already uses for non-200/error/timeout).

Implementation Decisions

• Thread a depth counter (or remaining-redirect budget) through the inner get(); default cap 5.
• On cap exceeded: res.resume() to drain, then resolve(null).
• Keep the existing null-on-failure contract — callers already treat null as "fetch failed".
• Guard against a missing location header (resolve null rather than get(undefined)).

Testing Decisions

• Add a node --test case in tests/install.test.js using a stub https.get that always returns a 301 to itself; assert the promise resolves null and does not hang/overflow.
• Add a case for a finite redirect chain within the cap resolving the body.

Out of Scope

• Changing RAW_BASE or the local-first fetch strategy.
• Retry/backoff on network errors.
• tasks.json slug-loading behavior (Wire Node installer to read task slugs from tasks.json #65).