eval: record real-provider benchmark numbers (M9 follow-up)

[div0rce]

Context

M9 (PR #12) shipped the evaluation harness + asserted-fixture pytest +
methodology-only PENDING eval/RESULTS.md. No numerical metric is committed
to the tree per CLAUDE.md Golden Rule #5 — quotable numbers come only from a
real-provider run.

This issue tracks running the harness against real providers and committing the
real eval/RESULTS.md plus a PROGRESS.md "Decision log" entry recording the
numbers, the run date, and the exact provider/model strings.

What "real-provider" means here (locked in by M9)

• LLM: claude-sonnet-4-6 (Anthropic). Verified against
  https://docs.anthropic.com/en/docs/about-claude/models on 2026-05-29 — the
  4.6-generation IDs use a dateless format that is itself a pinned snapshot,
  not an evergreen pointer.
• Embeddings: text-embedding-3-small (OpenAI), 1536 dim, matches
  backend.app.models.SCHEMA_EMBEDDING_DIM.
• Temperature: 0.0 per CLAUDE.md house style.
• k: 5 (default Settings.retrieval_top_k).

Recipe

Branch off main (e.g. eval/real-numbers-202Y-MM-DD):

export ANTHROPIC_API_KEY=...
export OPENAI_API_KEY=...
export LLM_PROVIDER=anthropic
export EMBEDDINGS_PROVIDER=openai
# (defaults already set CLAUDE_MODEL=claude-sonnet-4-6 and
#  OPENAI_EMBEDDING_MODEL=text-embedding-3-small in backend/app/config.py)

# Apply migrations to a fresh DB
make migrate

# Seed the synthetic corpus
make seed

# Run the harness — overwrites eval/RESULTS.md with real numbers
make eval

# Sanity check: nothing in the file should be a fabricated number
git diff eval/RESULTS.md

Then add a one-line entry to PROGRESS.md "Decision log":

- 2026-MM-DD (M9 follow-up) — Real-provider evaluation results recorded:
  extraction micro=X.XX macro=X.XX, retrieval p@5=X.XX r@5=X.XX MRR=X.XX,
  RAG citation-validity=X.XX cites-relevant=X.XX. Run with claude-sonnet-4-6
  + text-embedding-3-small at temperature=0; full breakdown in eval/RESULTS.md.

Open a small PR titled eval: record real-provider numbers (or similar; this
is a short, single-purpose PR — no other changes).

Acceptance criteria

• eval/RESULTS.md has been overwritten by make eval running against
  real LLM_PROVIDER=anthropic and EMBEDDINGS_PROVIDER=openai.
• Every committed metric value matches the harness output exactly. No
  hand-edits to numbers.
• The "Run metadata" section reflects the actual provider, model, embedding
  model, dim, temperature, k, and run timestamp.
• PROGRESS.md "Decision log" gains one entry summarising the headline
  numbers + the provider/model pair + the date.
• CI passes (existing backend + frontend jobs both green; no eval-job
  changes needed).
• No API keys committed (use local env or, for an actual remote run, GitHub
  Actions secrets — not done by default for cost control).

Out of scope

• Expanding the labeled set beyond the M9 5-invoice / 6-query / 5-question
  shape (separate backlog item, post-M11).
• Adding an LLM-judge faithfulness metric or ROUGE/BLEU (correctly excluded
  in M9; revisit only if the dataset grows).
• Running the eval in CI on every PR (would burn API tokens; stays manual or
  on a manual-dispatch GitHub Actions workflow if and when one is added).

References

• M9 PR (harness contract): feat: evaluation harness and benchmark results #12
• Methodology defense: docs/evaluation.md
• Asserted scorer/writer behaviour: backend/tests/test_eval_harness.py
• CLAUDE.md Golden Rule feat: schema-constrained structured extraction #5: "Never fabricate evaluation numbers."