hi-eval: per-turn context-growth harness + long interp task

[davidrhodus]

Why

The Pi-efficiency edge (send less context, hold/raise pass rate) lives in long multi-turn sessions where context accumulates — the exact regime the single-turn bench can't see. The eval already drives multi-turn (HI_EVAL_GOAL=1 HI_EVAL_TURNS=N, session persisted via --session-file), but it overwrote the report each turn, so only the last turn's tokens survived. No growth curve → the regime stayed unmeasured. This adds the missing instrumentation.

What

Per-turn growth capture. A new TurnMetric is recorded after every session-continuing turn — context (input_tokens) sent, total tokens, tool calls, and goal progress (done/total, read from the report's goal summary block). Threaded Candidate → RunResult → RunArtifact (serialized into each trial JSON + runs.jsonl) and printed as a per-turn curve:

growth (ctx tok · tools · sub-goals done):
  t1  ctx=    9450 tools=  3 goal=0/8
  t2  ctx=   28407 tools=  2 goal=8/8

input_tokens/turn is now a first-class signal — the "how fast does context accumulate vs. how far does the goal get" axis. Watch for ctx climbing while done stalls: that's the bloat the curation/compaction levers should flatten.

A long task — bench/long/interp: a ~10-requirement expression interpreter (tokenizer + recursive-descent parser: precedence, right-assoc **, unary, variables, comparisons, min/max/abs, typed errors) with a 30-assertion offline verify and a tested reference. Joins the existing multi-file CLI builds (kanban-cli, library-cli).

Verified live (pipe/glm-5.2-fast, goal-driven multi-turn)

HI_EVAL_GOAL=1 HI_EVAL_TURNS=8 HI_EVAL_TURN_STEPS=3:

• ctx 9,450 (t1, goal 0/8) → 28,407 (t2, goal 8/8), PASS — context accumulation and goal progress both captured and displayed.
• A second run (TURN_STEPS=8) independently showed 38,424 → 54,713 across two turns.

Honest caveat: interp is a single-file task, so a capable model writes it in one big edit and completes in ~1–2 turns — it demonstrates the curve but won't produce a many-turn one. The natural long-curve demonstrators are the multi-file kanban-cli/library-cli tasks; running those multi-turn (and then A/B-ing curation/elision/compaction levers against the curve) is the follow-up, gated on stable pipenetwork capacity.

How to run

HI_API_KEY=… HI_BIN=…/target/debug/hi \
  HI_EVAL_GOAL=1 HI_EVAL_TURNS=8 HI_EVAL_TURN_STEPS=3 HI_PLANNER_MODEL=pipe/glm-5.2-fast \
  cargo run -p hi-eval -- --profile=pipenetwork --models=pipe/glm-5.2-fast --configs=baseline --trials=1 bench/long

Test

• New read_turn_metric unit test (real report goal.done/goal.total shape + fail-open on a missing report).
• cargo fmt --all --check ✓ · cargo clippy -p hi-eval -D warnings ✓ · 14 hi-eval tests pass.
• interp passes --validate (fail-before / pass-after); data-only task files, no CI surface.

🤖 Generated with Claude Code