test: add live end-to-end validation harness

[JNK234]

Stacked on #74. Targets feat/structured-outputs-22 because T6 exercises llm:chat-with-schema, llm:chat-json and llm:get. Merge #74 first and this retargets to main cleanly.

Why

The unit suite is deterministic by design — it asserts request payloads without a live call, so CI needs no keys. That proves the plumbing, not that a provider accepts the request or that its response parses.

Today's review cycle showed the gap is not theoretical. Two findings derived carefully from vendor documentation were each settled in the opposite direction by one real call:

Claim from doc review	What the live call showed
Groq defaults to reasoning_format: "raw", so <think> tags leak into visible content — rated HIGH	False for gpt-oss-20b. Returns clean content plus a separate reasoning field.
reasoning_effort mismatch — rated MEDIUM	True and worse. xhigh passed extension validation and returned HTTP 400.

What

demos/e2e-tests/ — six test procedures making real API calls through the installed extension, with a chooser for the provider under test.

Test	Covers
T1	plain chat round trip
T2	thinking-tag leakage in visible content
T3	thinking enabled — reasoning arrives separately, not duplicated
T4	llm:compile-error (#52)
T5	per-agent history isolation and clear-history
T6	structured output (#22)

T6 asserts types, not just shape. A provider can return well-formed JSON and still fail if the conversion drops types, so is-number? and is-boolean? are checked alongside the enum constraint, and a missing key must raise rather than default silently.

Verification

22/22 pass against live Groq (openai/gpt-oss-20b), headless, real HTTP, no stubs:

T6 structured output
  -> [[action eat] [confidence 0.9] [alive true]]
  PASS  action honours the schema enum
  PASS  confidence is a NUMBER, not text
  PASS  boolean field is a boolean
  PASS  missing key raises rather than defaulting
  chat-json -> {"colours":["red","blue"]}
TOTAL  passed 22   failed 0

config-ollama.txt needs no key, so the suite runs offline against a local model too.

Keys

config-groq.txt and config-anthropic.txt are gitignored and ship as .example templates. Only config-ollama.txt, which needs no key, is tracked.

Headless finding worth recording

NetLogo 7.0.3 only runs an experiment embedded in the .nlogox. An external --setup-file is accepted and then silently ignored, aborting with NoSuchElementException: head of empty list before setup runs — indistinguishable from the benign 0-step abort. The embedded block additionally requires sequentialRunOrder, timeLimit as an attribute rather than a child node, and metrics wrapped in <metrics>.

This cost real debugging time and is not documented anywhere in the repo.