Supersedes v0.6.0-anthropic (which carried the retracted "0/180"
walker-bug artifact — do not cite that tag).

Corpora:
- v0.5 OpenAI long-plan (re-scored with fixed walker)
- v0.6 Anthropic long-plan (Haiku 4.5, Sonnet 4.6, Opus 4.7)
- v0.6 short-plan (5 scenarios: travel maintenance, peaking,
  postpartum on-ramp, post-illness reconditioning, deload)

Headline findings (see docs/V0_6_RESULTS.md):
- The WPL contract reduces unsafe-trial rate 3-5x across every corpus
  and both phases (single + multi-turn) — not to zero, as the buggy
  v0.6.0-anthropic claimed
- Lane A safety degrades with model capability: the two flagships
  (Opus 4.7: 129, gpt-5: 89 combined long-plan violations) are the
  two worst raw performers; the cheapest (Haiku: 27, gpt-4.1: 29) are
  safest
- The schema validator, not the compile gate, is the served-rate
  ceiling; failures concentrate at the activity block
- DSL->compile does real schema-conformance work direct JSON emission
  can't substitute for (native-JSON probes)
- Short-plan corpus surfaces structural failures (rest days,
  progression, on-ramp) the raw-LLM prose lane is blind to; in
  multi-turn the conversation drives structural drift the contract
  surfaces but doesn't yet prevent

Four bugs found and fixed since v0.6.0-anthropic: the Lane B JSON
walker, four short-plan scorer rules, multi-turn final-turn semantics,
and markdown-fence stripping. Full disclosure in the RESULTS doc.

560 trials, ~$170 total inference. Fully reproducible from stored
output without re-running the LLM.