SMA: add bare-model stumper triage harness

[seanoc5]

SMA: add bare-model stumper triage harness

Adds tooling under tests/evals/triage/ to identify questions a bare
LLM cannot answer correctly about Solr→OpenSearch migration — the
foundation for deciding which skill content matters.

What it does

• 20 candidate questions in tests/evals/stumper-candidates.md, each
  with two-criterion contains-any assertions. Organized by topic:
  Querqy/SMUI, schema migration, query translation, AWS-managed
  constraints, consultant judgment.
• tests/evals/triage/sweep.py runs each question against Claude
  Haiku 4.5 (cloud, via claude --print from /tmp for skill
  isolation) and qwen2.5:7b (local Ollama HTTP). Responses cached by
  sha256(model + prompt) so iterating on criteria is free.
• tests/evals/triage/responses/*.txt — bare-model responses for all
  20 questions × 2 models (42 files including q05's pre- and
  post-reframe variants). Reviewers can see the failure mode without
  re-running.
• tests/evals/triage/triage-results.md — pass/fail summary table.

What this surfaces

Of 20 questions, 12 are stumpers (both models fail at least one
criterion). The 7-question subset used by the follow-up eval
(#49) all originate from this triage.

Test plan

(See first author comment for full run-through.)

[seanoc5]

Test plan

Verifies the harness runs and emits responses + cache + summary table.

cd AIAdvisor/skills/solr-opensearch-migration-advisor/tests/evals/triage

# Prereqs
which claude    # Claude Code CLI (uses local OAuth credential)
ollama list | grep qwen2.5:7b   # qwen2.5:7b pulled locally

# Smoke test: one question, both models
python3 sweep.py q01-querqy-class
# Expected: ~10s total. responses/q01-querqy-class.{haiku,qwen}.txt
# created. triage-results.md updated with one row showing A=RED B=RED
# for both models.

# Cache test: re-run, should be 0.0s per call
python3 sweep.py q01-querqy-class
# Expected: "(cache)" tag in both lines, 0.0s timing.

# Cache-skip test
python3 sweep.py --refresh-model haiku q01-querqy-class
# Expected: haiku "(live)" ~7s, qwen "(cache)" 0.0s.

# Full 20-question sweep (~10 minutes)
python3 sweep.py
# Expected: triage-results.md table populated for all 20 questions.

Smoke test results from my own run are committed at
responses/q01-querqy-class.{haiku,qwen}.txt — Haiku honestly
declines, Qwen guesses the wrong package
(org.opensearch.queries.rewrite.commonrules.CommonRulesRewriter),
both A=RED.

Notes

• cache/ and *.json are .gitignored — addressable by hash, not
  useful in diffs.
• The claude CLI uses the local ~/.claude/.credentials.json OAuth
  credential; no ANTHROPIC_API_KEY env var needed.
• Ollama provider hits http://localhost:11434/api/generate by
  default.