v1.5.0

• Twelfth Scene fallacy lens (#44). Add fallacy_gamblers_fallacy
  with a history-dependence test that distinguishes unsupported local
  compensation in an unchanged independent process from real depletion,
  negative dependence, state change, parameter learning, future-block
  probability, and regression to the mean. Unknown generator rules return
  insufficient_context; unsupported continuation raises a distinct off-list
  hot-hand question. The per-lens not_fallacy ruling now explicitly avoids
  certifying the whole argument. Gate 10M, invariant/mutation coverage, and
  artifact-carrier checks carry the contract without a Passport schema or
  taxonomy change.
• Eleventh Scene fallacy lens (#42). Add fallacy_motte_and_bailey
  with a fallback-substitution test that requires a stronger claim, a narrower
  fallback, and observable evidence that the stronger claim or its conclusion
  remains in force. Explicit narrowing or withdrawal is legitimate when the
  stronger conclusion is abandoned, materially weakened, or independently
  supported without substituting the fallback; a later unsupported return still
  triggers the lens. Incomplete passages return insufficient_context.
  Nearby-lens boundaries plus no-motive and accountable-advocate guards prevent
  overreach. Gate 10L and invariant/mutation coverage carry the contract without
  a Passport schema change.
• Concurrent local Passport checkpoints (#38). Route every Claude Code
  checkpoint and deletion through one bundled writer with an exclusive sidecar
  lock, lock-before-reread ordering, whole-batch atomic replacement, and
  fail-closed pending-event retention. A deletion-generation token prevents old
  sessions from restoring pre-deletion batches; strict JSON validation,
  no-follow regular-file checks, and locked orphan-temp cleanup harden the local
  privacy boundary. Real multi-process tests cover simultaneous writers, cold
  start, malformed input and tails, lock contention, deletion and SIGKILL
  recovery, permissions, and failed replacement without changing the event
  schema. The local writer explicitly requires Node.js 22+; its preflight exits
  before touching Passport data when that runtime is unavailable, while the
  coaching session continues with recording paused.
• Source credibility becomes the 14th canonical structure (#36). Add
  source_credibility as a loggable Drill and Detective target while retaining
  the existing cross-mode clarify, check_basis, and
  license_conclusion as procedures rather than duplicate tally IDs. Existing
  drill_result, miss_log, item_discarded, and
  detective_process.structures_hit events carry it without a Passport schema
  bump or a fifth Drill item type.
• Credibility-specific generation and behavior gates. Require uniquely
  defensible weight judgments across tiers, including sound interested sources;
  refuse to invent motives, funding, independence, credentials, or records;
  preserve the genetic- and ad-hominem-fallacy reverse guards; and protect
  Detective's first-defect-call and answer-chain silence. New invariants,
  mutation coverage, eval templates, and manual probes keep those boundaries
  load-bearing.
• Source-family answer-cue repair (#35). Rebalance the seven public
  source-credibility Daily cases so the correct option is no longer revealed by
  conspicuously greater length, tighten the 2026-07-19 prompt, and add a
  keyless regression test for the length cue. Private answer records remain
  local and outside version control.
• Scene steelman ordering. Require the coach to reconstruct a caricatured
  position as a complete claim with its strongest stated reason before
  correcting the caricature; merely listing objection fragments no longer
  satisfies the steelman duty. An invariant and mutation check keep the
  ordering load-bearing.