v0.7.0

Added

• Two production-shaped examples alongside BDR, each adapted from a
  real production skill with all identifiers fictionalized:
  examples/ops-report/ (the 100%-roteness archetype — every step
  deterministic, one durable HITL gate, zero LLM nodes; the fixture for
  the python adapter's durable-execution refusal) and
  examples/deal-monitor/ (the data-heavy archetype — parallel entry
  waves, fan-out judges, template render replacing LLM-generated HTML;
  the calibration fixture for the payload-aware estimator).
  tests/test_examples.py guards every example's expected IR, including
  that its source_skill pointer resolves.
• Payload-aware "before" cost estimator — the static scorecard now
  models the data a skill pulls into context, not just its turn count.
  The compiled pipeline's external_call footprint sizes the agent-side
  context payload (tokens_per_external_call_result, default 6k/call,
  with a per-MCP-tool override table payload_tokens_per_tool), folded
  into C₀ of the cache-aware transcript model. Calibrated against a real
  data-heavy production skill (Slack + Gmail dashboard, ~22 turns,
  ~1.6M cache-read tokens/run): the old flat prior underestimated the
  before-cost 5–15×; the payload-aware default lands within ~3×, and one
  eval --run calibration brings it within ~10%.
• rote eval --run's suggested prior re-fits now include
  transcript_growth_per_turn, inverted from measured cache-read tokens
  under the quadratic transcript model — so every empirical run reports
  the effective payload-inclusive growth rate alongside
  seconds_per_turn and output_tokens_per_turn.
• CodexDriver is now implemented — rote compile --agent codex
  spawns codex exec (OpenAI Codex CLI) as a compiler backend,
  completing the three-driver lineup. Runs headless under a
  workspace-write sandbox (global reads so it can read the skill +
  rubric in place, writes confined to the work dir, no network). The
  environment is passed through untouched — a stored codex login
  session is only overridden by CODEX_API_KEY/CODEX_ACCESS_TOKEN,
  not OPENAI_API_KEY, and there is no separate OpenAI-API driver, so
  no auth is forced. Verified against the real CLI (codex-cli 0.142.4).
• rote analyze is now implemented — the plan to compile's
  apply. Runs the compiler against a skill and prints a structural
  report (node-kind breakdown, roteness — matching rote eval — plus
  mandatory checks, HITL gates, agent loops, and which runtimes can
  target it) without emitting runtime code. --json for a
  machine-readable report; --out keeps the compiled IR for a later
  rote emit. Previously a stub that printed "not yet implemented".

Changed

• source_skill no longer participates in the pipeline hash — it's
  provenance (a filesystem path the compiler re-points per output
  location), and hashing it minted a new workflow type on every
  re-compilation to a different directory, re-versioning in-flight
  workflows whose behavior didn't change. Same rule as Node.source,
  which was already excluded. One-time consequence: every pipeline's
  hash (and therefore emitted workflow type name) changes once with this
  release; in-flight workflows on the old type names continue on old
  code as designed.

Fixed

• The codex driver no longer raises NotImplementedError when
  selected via --agent codex or chosen by auto-detect (a first-run
  crash for users who had the Codex CLI but not Claude Code installed).
• The BDR example's committed IR baseline carried the same dead
  source_skill pointer the orchestrator fix addresses
  (../../skill resolved to a nonexistent examples/skill), so
  rote eval on the canonical example silently dropped its before-side
  baseline. Corrected to ../skill; now regression-guarded for every
  example.
• rote compile/rote analyze --out now re-point the pipeline's
  source_skill to resolve from the emitted pipeline.yaml's location
  (relative when possible, absolute otherwise). The agent records the
  path relative to its temp work dir — deleted when the run ends — so
  every kept compilation carried a dead pointer and a later rote eval
  silently dropped the entire before-side baseline ("source_skill did
  not resolve — emitting the after-side only"). Found by compiling a
  real production skill.