v0.8.0

Added

• examples/invoice-push/ — the fourth committed example and the
  agent_loop archetype, completing node-kind coverage across the
  examples (adapted from a real production browser-automation skill,
  fully fictionalized): a bounded per-row browser cycle (5-node
  loop_body, termination cap) that cannot crystallize into code,
  surrounded by 12 deterministic nodes. First example to commit its
  eval.yaml sidecar — with per-row iterations, the calibration
  fixture for the loop-aware cost model below.

• eval --run auto-wires the pipeline's MCP servers into the skill
  trial. The compiled pipeline's mcp: bindings name exactly the
  servers the source skill uses, so the "before" measurement now runs
  the agent over the same live tools (--mcp-config +
  --strict-mcp-config, per-server mcp__<server>__* allowlists;
  URLs resolve from the binding's explicit url, else
  ROTE_MCP_<SERVER>_URL — the adapters' rule). Verified live against
  a real claude -p run over a Streamable-HTTP MCP server.

• Reliability flags on measured skill runs. Each trial is checked
  structurally (never by an LLM) before it may calibrate priors:
  errored, hit_max_turns (truncated — its cost is a floor, not a
  measurement), suspiciously_few_turns (fewer turns than the pipeline
  has data pulls — the agent never did the work), and
  missing_mcp_servers (the skill ran without its tools). Flagged runs
  are excluded from suggested_priors re-fits, listed in the measured
  scorecard section, and recorded in the calibration corpus with their
  flags.

• Loop-aware before-cost model, calibrated against a production
  browser-automation skill whose two real runs measured 184 and 730
  agent turns (~$8.6 and ~$32 on Sonnet) against a prior estimate of
  $0.82–$2.75:

  • The eval sidecar's StepEstimate gains iterations: {low, high} —
    a step that repeats per row/page/item declares its per-iteration
    turns and realistic repeat count, and the whole-run turn estimate
    multiplies them. Loop-dominated skills were previously understated
    5–20× because the schema could not express iteration at all (the
    compiler's own notes described the loop economics correctly; the
    form had nowhere to put the number).
  • New transcript_cap_tokens prior (default 165k, the per-turn
    cache-read plateau measured on both production runs): the transcript
    an agent re-reads saturates at the harness's compaction ceiling
    instead of growing without bound, so cached-read totals transition
    from quadratic to linear on long runs. Without the cap, correcting
    the turn count would have swung the error the other way (~2.5×
    over).
  • The compiler rubric's calibration anchors are now regime-aware:
    sequential tool-heavy skills anchor at 30–57 turns, per-item loop
    skills at hundreds (the old universal 30–57 anchor actively pulled
    loop-skill estimates down an order of magnitude).

  Re-estimated with both fixes, the calibration skill's scorecard
  brackets reality: $1.96–$22.68 estimated vs $8.59–$32.00 measured
  (Sonnet), 98M estimated token ceiling vs 96M measured.

Fixed

• rote compile now re-points eval.yaml's source_skill alongside
  pipeline.yaml's — the sidecar previously kept the agent's
  temp-work-dir-relative path, a dead pointer in every kept compilation.