feat: add llm:compile-error for validating LLM-generated NetLogo code

[JNK234]

Closes #52 (roadmap B3).

Why

Several demos use the "LLM writes NetLogo code, the model runs it" pattern (demos/templates/code-evolution-template.yaml, movement-evolution.yaml). Today generated code goes straight to run/runresult with no checks: invalid code throws mid-tick, and nothing stops a model from emitting clear-all or die.

This adds the gate that makes that pattern safe enough to put in front of students.

What

A single reporter, llm:compile-error, returning "" when code is acceptable and an explanatory message otherwise.

llm:compile-error "fd 1"        ;=> ""
llm:compile-error "frobnicate"  ;=> "Nothing named FROBNICATE has been defined. (offset 0)"

(llm:compile-error code ["die" "clear-all"])
;=> "Disallowed primitive(s) used: die"

Design decisions

Uses NetLogo's own compiler, reached via ExtensionContext → workspace.compileCommands. This answers the issue's first open question: an extension can compile arbitrary source without a full workspace round-trip. It also means the live model's symbol table is in scope, so turtles-own variables, globals, and breeds resolve — an external validator cannot do this, and undefined-variable errors are the most common failure caught.

Compiles in turtle context, because generated agent rules are normally executed inside ask turtles. Observer context would reject fd/rt and refuse valid rules.

Syntax is checked first and returns early. Code that does not compile cannot run, so reporting banned primitives on it would be noise — and tokenizing malformed code is unreliable.

Banned-primitive matching is on exact tokens via NetLogo's tokenizer, not substrings. Banning die rejects die but not diehard, not die in a comment, and not the string "die". Substring matching produced all three false positives, and a false positive discards code that was valid.

No built-in banned set ships. What counts as dangerous is model-specific — a model may legitimately need hatch while forbidding it in generated rules. The issue suggested a default list; this was deliberately not taken, so policy stays in NetLogo where it can change without rebuilding the extension.

A malformed list is rejected rather than ignored. (llm:compile-error "die" [1 2]) previously returned "" — a false all-clear for a caller who passed the wrong thing. It now raises.

Scope

This proves code compiles, not that it cannot throw at runtime. Bounds errors such as item 3 on a three-element list still surface only during execution, so carefully around run is still needed. Documented in the API reference.

Tests

sbt test — 54 passing, 0 failed. 15 new blocks in tests.txt (~70 assertions):

• valid / invalid code, turtle context, multi-line let scoping, comments, nested blocks
• a 14-rule corpus of real evolved rules from the LEAR project, as a false-negative regression guard
• banned primitives: hit/miss, case-insensitivity both directions, dedup, multi-line
• the three false-positive cases (comment, identifier, string literal)
• ordering (syntax error wins over banned report) and malformed-list rejection

Follow-ups, not in this PR

• A demo model exercising the primitive
• The default-banned-list question, if a sane default turns out to be wanted after real use