docs: six runnable examples with captured output

[queso]

What

Adds examples/: six runnable, self-contained example directories ordered as a learning curve, each with its own agent, skills, scenario file, README, cost label, and captured output.

Example	Shows	Cost
01-hello-compare	Smallest baseline-vs-proposed compare: one added instruction, one text grader, 0% to 100%	~$0.01
02-fix-a-regression	Target + regression pair on commit-message subject length	~$0.10
03-measure-first	Single-arm measure on an under-specified extraction prompt	~$0.01
04-judge-grader	Calibrated LLM judge end to end: rubric, fixtures, calibrate, then compare	~$0.10
05-any-model	openai runner against local ollama, no API key	free
06-token-economics	--max-turns and --raw-out for token-category analysis	~$0.19

The root README gains a "See it work" section pointing at 01.

Why the test is in here

test/examples.test.ts loads every examples/*/scenario.json through the real loadCompareConfig, including its file-existence validation. Without it the examples are documentation that drifts from the schema and breaks for the first user who copies one. With it, drift breaks CI instead.

Verification

• bun run check: 120 tests pass, typecheck clean.
• audience-appropriate.md hashes to the rubricSha256 in its committed calibration record, and assertJudgeCalibrated passes on the real grader spec, so 04 does not refuse on a fresh clone.
• Every output string pasted into a README matches its source: the grader messages (grader.ts:72, :84), the judge verdict format (judge.ts:282), the sampling-noise NOTE (compare.ts:157), the turn-cap failure (compare.ts:472), the calibration refusal (judge.ts:198), and the raw filename (compare.ts:529).
• 02's ^.{1,50}(\n|$) grader is correct because gradeText compiles patterns with no m flag, so it anchors to the start of the output rather than to any body line.
• 06's seed contains no occurrence of "backoff" or "exponential", so the task cannot be grepped and queue.py is unambiguous among the five files.

Doc accuracy fixes (2d9d5d4)

Three inaccuracies the first commit shipped with, plus a fourth of the same
kind found while checking them:

• examples/README.md and the root README claimed every example's output was
  captured from a real run. 05 was never executed (no ollama in the build
  environment) and its own README says so. Both now name 05 as the
  exception, and 05's table row is labeled illustrative.
• The root README's "See it work" block called itself real output while
  omitting the NOTE: delta could be sampling noise line and the two
  baseline run N failed lines that a real run of 01 prints. It now shows
  01's full output.
• 05's illustrative block claimed to be shaped exactly like
  formatCompareSummary's output but omitted the per-failing-run lines a 1/3
  baseline emits (compare.ts:161-168 pushes one per failing run
  unconditionally). Added.
• 06's README did not state that its grader is a substring check, so a
  reader could take the 2/2 pass rate as measuring the single-line format the
  prompt and skill.md ask for. Both committed raw results show a sentence of
  reasoning ahead of the filename and still passed. Now stated, with the
  {"type": "text", "regex": ["^queue\\.py\\s*$"]} grader that would
  enforce the format (regex, not matches — GraderSpec has no matches
  field).

bun run check still green: 120 tests pass, typecheck clean. Markdown-only
changes; no example was re-run, so no captured number moved.

🤖 Generated with Claude Code

[github-actions]

Nitpick review — approve

The change adds documentation and worked examples under examples/ with a small test that loads each scenario to catch schema drift in CI. The examples are additive and don't touch production code paths, and the accompanying test exercises every scenario's load path. No issues that warrant changes were identified.

0 inline comment(s).

Verdict was approve, posted as a comment — this repo does not allow GitHub Actions to approve PRs.

[github-actions]

Nitpick review — comment

The change adds documentation and worked examples for the eval tool, with no production code touched. One arithmetic inconsistency survives in the README's captured-output block: the delta line reads +$0.0013 while the baseline and proposed costs ($0.0021 and $0.0033) imply +$0.0012, and the same block appears in examples/01-hello-compare/README.md. Worth correcting so the docs stay internally consistent; otherwise the change is sound.

1 inline comment(s).

[github-actions]

suggestion · Arithmetic error in documented sample output: +$0.0013 should be +$0.0012

The captured output block shows baseline $0.0021 and proposed $0.0033, so the delta line should read +$0.0012, not +$0.0013. The same block is duplicated in examples/01-hello-compare/README.md (line 24). Since the README explicitly claims the output is 'captured from a real run', an internally inconsistent cost figure undermines that claim.

Suggested change

	delta: +100% pass | +$0.0013
	Correct the delta to +$0.0012 in both README.md and examples/01-hello-compare/README.md, or re-capture the block from an actual run so the numbers are self-consistent.