llmtrim 0.1.6

Added

• Range-fold for regular sequences in tool-output template folds: when a folded
  log's parameter column is a regular sequence — constant values, arithmetic integers,
  or constant-step ISO-8601-like timestamps — the explicit value list collapses to a
  lossless range ([×30: (10:02:00Z..10:02:29Z step 1s; 0..29)]). Every value stays
  byte-exactly reconstructible (a round-trip check gates each fold); irregular columns
  keep the explicit list, and a range is emitted only when strictly shorter. On the
  README's build-log example the same request now compresses −71% instead of −62%.
• Missed-fold telemetry in the capture loop: with LLMTRIM_CAPTURE_DIR set,
  datetime-ish columns that fall back to the explicit list are logged to
  missed_folds.jsonl (reason + 5-value sample), so real traffic — not guesswork —
  decides which timestamp shapes the range fold learns next. Zero overhead when
  capture is off; a write failure can never break a fold.

Fixed

• Re-run → passthrough rail now survives non-deterministic output: the rail that
  ships a re-invoked tool's output in full used raw-text equality, so any run-to-run
  noise — TAP's duration_ms timings, log timestamps, ports, PIDs — defeated it and
  the retry was windowed identically. Repeat detection now compares a
  volatile-value-masked fingerprint (the template stage's variable masking), so a
  re-run that differs only in such values passes through in full, while a real result
  change (a test flipping ok ↔ not ok) still compresses fresh.
• TAP test failures no longer elided (reported against v0.1.5): a node --test /
  prove TAP log could lose its only failing test — not ok N, the YAML diagnostic,
  even the # fail 1 summary — because the failure-signal regex didn't know TAP's
  not ok marker (nor camelCase tokens like failureType: 'testCodeFailure'), and
  the retrieve stage ranked chunks purely by query relevance with no failure
  protection at all. Failure-signal lines and their continuation blocks (indented
  traceback frames; for TAP, the whole diagnostic up to the next test point) now
  survive pruning in both the tool-output and retrieve stages, regardless of query
  overlap.