Eval: /goal plan-completion fixture (measure % of plan delivered)

[davidrhodus]

The MVP goal eval — a number for "does /goal finish a plan?"

Measures what /goal is actually for: given a multi-section plan.md and "build all of it", what fraction of the plan does the agent deliver? The hidden oracle scores per section and emits a fraction, so a run that finishes 4 of 6 sections scores 0.67 — not a flat fail. That's what turns #107 from "should help" into a measured before/after (e.g. the qtest reality of ~20% → whatever the fixes actually achieve).

What's here

• bench/plan/textkit/ — six independent, turn-sized Python modules (slugify, wordcount, roman, caesar, rpn, csvmini), each an exact contract with examples in fixture/plan.md. The hidden oracle/score.py checks each section and prints sections delivered: X/6 + HI_EVAL_SCORE=<0..1>; fixed/ is a reference that scores 100%.
• bench/plan/README.md — how to run it (goal-drive mode, baseline vs candidate) and read the score.

Reuses the existing harness — no harness code changed

• Runs via the existing hi-eval goal-drive mode: HI_EVAL_GOAL=1 HI_EVAL_TURNS=<n> drives the goal across session-continuing turns (the real /goal cadence).
• The fraction rides in the oracle's captured output; a fully-delivered plan also exits 0 (a normal harness pass).

Validated (deterministic, no model)

cargo run -p hi-eval -- --validate bench/plan passes — the empty fixture/ scores below 100% (fail-before), fixed/ scores 100% (pass-after), the oracle is un-gameable by candidate-created files, and allowed_changes is sane. I also ran the scorer directly: fixed/ → 6/6 (100%, exit 0); dropping two modules → 4/6 (67%, exit 1) with no crash.

How to use it

HI_BIN=<hi binary> HI_MODEL=<provider/model> HI_API_KEY=<key> \
HI_EVAL_GOAL=1 HI_EVAL_TURNS=12 \
cargo run -p hi-eval -- bench/plan

Run against the old binary and a #107 build; compare the HI_EVAL_SCORE fractions.

Follow-ups (not in this MVP)

• Surface HI_EVAL_SCORE as a first-class aggregated metric in the harness report (today it rides in the oracle artifact).
• Add larger / graded plan fixtures (medium, large) once this shape is agreed.

Note: this branch is off main, independent of the goal-engine PR (#107) — the fixture runs whatever hi binary you point it at, so it measures either.