stratless 0.3.0

The learning release. Until now the profile was re-derived from raw judgment lines on every build
and the derivation thrown away; nothing accumulated. 0.3.0 makes the derivation a persistent,
auditable artifact — the profile stops being an impression and becomes a derivation from evidence.

Added

• The miner. A new pass turns your judgments into ~/.stratless/patterns.json: named
  regularities, each carrying a real count, a time window, a trend (rising/steady/fading), a
  stability class, and — the part that matters — receipts: the exchange hashes that witnessed
  it. Every claim traces to replies that could have gone the other way; no receipt, no claim. The
  split is strict: the model names, the code counts — counts, dates and trends are computed
  deterministically and the model is never shown a number it could hallucinate. Categories emerge
  from your data (a pattern needs 5+ receipts to exist; anecdotes stay candidates); they sort into
  six fixed kinds — what you know · how you think · how you work · direction · failure signals ·
  triggers — and anything that fits none is kept unsorted, never forced.
• stratless patterns — every claim with its count, trend, audit tally, and receipt trail.
  Your profile's evidence, inspectable. Free; add --all for candidates below the evidence bar.
• The auditor. A separate mind — never the one that wrote a statement — re-checks every newly
  assigned receipt against its pattern ("does this judgment actually SATISFY the statement?").
  Evicted evidence is removed by code and admission is re-decided by arithmetic. Audited once,
  ever, per receipt.
• The numbers-lint. Every numeral in a built profile must already exist in the evidence the
  writer was shown. An invented or rounded frequency is a refused build — the old profile stays
  loaded, nothing lies. A wrong frequency is a lie wearing precision; now it's a build failure.
• Dynamic aperture. The judge's view of each exchange is sized from your own history (p90 of
  your real lengths × 1.2, clamped), reading head + tail of long turns — the plan and the
  conclusion — instead of a fixed head-cut sized on the author's machine. Recorded in
  state.json, visible.

Changed

• Judgments are structured (v2). Verdict / topic / behavior fields, JSON output validated in
  code (malformed = refused, never guessed). The pipeline version now lives in the cache key, so
  this release re-judges your window gradually, under the normal per-run budget — an upgrade
  never re-spends your backlog at once.
• The profile is written FROM the patterns — audited claims with real numbers, plus a small
  recent sample for freshness. The writer's input stays the same size whether you have 200
  judgments or 20,000: the learning architecture is the cost-control architecture.
• The parse cap rose 4,000 → 8,000 chars per field (inside the v2 re-judge, so the identity
  churn was free). Long assistant turns keep more of their real tail.

---

Published to npm over Trusted Publishing (OIDC, no token) with Sigstore provenance — verify with npm audit signatures.