feat: named profiles so different agents call different models

[JNK234]

Closes #68.

What

Provider and model were one global configuration shared by every agent, and swapping it per agent broke under async calls (an in-flight request resolved against whatever config the next agent installed). This adds named profiles: a second config file loaded under a name, with its own provider, key, model, thinking, retry and throttling settings. Any agent can be bound to a profile; unbound agents use the global configuration exactly as before.

llm:load-config "config.txt"
llm:load-profile "llama"  "config-groq.txt"
llm:load-profile "sonnet" "config-anthropic.txt"
ask turtles with [who mod 2 = 0] [ llm:use-profile "llama" ]
ask turtles with [who mod 2 = 1] [ llm:use-profile "sonnet" ]

Primitives

• llm:load-profile name file: validated exactly like llm:load-config (unknown provider, missing key, unreachable local server). Reloading a name replaces it; a rejected reload leaves the old profile intact. "default" is reserved.
• llm:use-profile name: binds the calling agent. "default" unbinds. Unknown name errors and lists what is loaded.
• llm:profile, llm:profiles.
• llm:active and llm:config now report the calling agent's effective configuration.

Semantics

• Each profile caches one provider instance, created on first use (ProfileStore in src/main/config). An async call captures its provider at start, so rebinding an agent affects its next call, never a pending one.
• Profiles survive clear-all like the global config does; bindings are cleared with the agents.
• llm:set-model, llm:set-provider, llm:set-api-key and the thinking setters keep acting on the default configuration only. A profile is what its file says.
• Throttling already keys on provider and endpoint, so profiles on the same endpoint share one cap. History and usage stay per agent.

Tests

• ProfileStoreSpec: 11 tests. Load, sorted names, reserved and blank names, case-insensitive lookup, provider created once and reused, reload drops the cached provider, unknown name lists loaded ones, factory failure not cached, concurrent first use creates one instance.
• tests.txt: 14 cases with two tracked fixtures (demos/test-profile-a, -b) and a __TEST_ECHO_MODEL marker so a reply reveals which config served it. Covers default profile, bad provider, missing file, reserved name, unknown profile, two turtles on two models with the observer on the default, llm:active/llm:config per agent, return to default, async keeps its provider after rebinding, set-model leaves profiles alone, reload replaces, failed reload keeps the old profile, clear-all resets bindings and keeps profiles, history and usage per agent.
• Live: T10 in demos/e2e-tests derives a second profile from the active Groq config with openai/gpt-oss-120b, binds one turtle, and checks routing, reporting, a real reply, and usage credit. 41/41 pass.

Deterministic suite: 249 passing.

Docs

docs/API-REFERENCE.md: new Profiles section. docs/CONFIGURATION.md: profiles subsection under provider examples.

Review follow-up

A Codex (gpt-6-astra) review of today's diff produced six findings. Three were profile-specific and are fixed in this PR with tests first:

• Throttle gates were replaced on any settings change, so two profiles with different caps on one endpoint bypassed the cap. Gates now reconfigure in place and keep the in-flight count.
• Profiles were built on OpenAI defaults, so a model-less Anthropic profile asked for gpt-4o-mini. Profiles now load like llm:load-config.
• Await budgets came from the global config. Calls now use the calling agent's effective config; async calls capture it at launch.

Three predate this work and are filed: #80 (async completion after clear-all), #81 (llm:choose on models that reject json_schema), #82 (refusal payloads lose usage).

Deterministic suite: 254 passing. Live harness: 41/41 on Groq.