feat: add built-in fetch skill (read webpage text) - #132

Adds a built-in fetch skill: pre-imported async fetch(url=...) that GETs a page with httpx (redirects followed, browser-ish UA) and returns its cleaned text — scripts/styles/templates stripped, tags collapsed, entities unescaped — truncated to max_chars (default 20k). Non-UTF-8 pages decode safely (errors="replace"). Network/HTTP errors come back as short Error: ... strings (data for the agent, not a crash).

Pairs with the search skill: search finds URLs, fetch reads them. A neutral prompt line advertises the skill and its signature when enabled.

• src/rlm/skills/fetch.py — the skill
• src/rlm/skills/__init__.py — registers fetch in the built-ins
• src/rlm/prompt.py — FETCH_SKILL_PROMPT, emitted when the skill is enabled
• README.md — documents fetch
• tests: enable-stub + html_to_text markup stripping

Measured effect (paired A/B, redsearcher, 100 tasks/arm, same pin, skills-only diff, no sub-agents)

*capped = per-session 100 turns, 100k context, 8k/call, 10k tool-out (identical across arms).

Reading: no significant reward gain from the skill itself. Mechanism from mining the control arms: a strong code model simply substitutes — deepseek hand-rolled 12,086 page reads (requests.get/urllib/curl) across 100/100 control rollouts, so fetch is a convenience wrapper for a capability the model already exercises. Laguna shows a small positive trend (+0.042, n.s.) with modest adoption (29/100 rollouts, 391 calls). The value of fetch is therefore primarily hygiene (clean text instead of raw HTML in context, uniform error strings, one sanctioned path that a truncation policy can target) rather than measured task reward.

Failure taxonomy from the runs (follow-up material)

2,971 fetch calls (deepseek arm): 33% error rate, dominated by 404 (835× — 631 on en.wikipedia.org: the model guesses /wiki/Title URLs from memory), then 403 (71), 429 (34). 28% of successful reads hit the 20k max_chars cap. Suggested follow-ups:

1. On 404, append a hint ("page not found — use search() to find the real URL") or auto-fallback to the site's search endpoint for known hosts (Wikipedia opensearch).
2. Continuation support (fetch(url, start_char=N) or a trailing "N more chars available" note) so truncated reads can page through.
3. Respect Retry-After on 429.

Also planned per review: expose fetch as a tool as well as a skill.

First of a 4-PR stack: fetch (this) → contract-only config (#160) → execution guardrails (#158) → role-aware sub-agent prompting (#151).