v0.7.1

Fixed

• The Log tab no longer renders whole CLI sessions underlined. Modern CLIs emit an XTMODKEYS
  sequence (CSI > 4 ; 2 m, "modifyOtherKeys") at startup that the pane emulator (pyte) misread as
  SGR 4 / underline-on — with no matching off present in a live capture — so every line came out
  underlined and hard to read. The log view now strips private-marker CSI sequences before emulation;
  genuine color, bold, and properly-closed underline styling is preserved.

• Resolving a CRITICAL escalation no longer loops on a manual-rollback prompt when the resolve
  edited the spec. 0.7.0 fixed the loop only for an already-clean tree, but the resolve workflow's
  whole job is to correct the frozen spec under the BMAD artifact folder (_bmad-output/..., which is
  tracked). So on resume the orchestrator saw a dirty tree and — with the default
  scm.rollback_on_failure = false — paused for a manual reset; because the dirty check diffs against
  the frozen baseline_commit, even committing the spec re-paused on the next resume, an endless loop.
  A resolved re-drive is human-initiated, so it now always auto-recovers regardless of the flag: the
  BMAD artifact folders are treated as orchestrator-owned — excluded from the dirty check and preserved
  through every reset of the re-drive (not just the resume-time cleanup) — so the spec correction
  survives while the failed attempt's source changes revert to baseline. This closes a latent sibling
  bug: with rollback_on_failure = true a later mid-re-drive retry/defer reset previously ran with no
  preserve set and reverted the just-corrected spec silently, looping the re-drive.
  scm.rollback_on_failure still defaults OFF and now governs only unattended/stopped attempts; the
  manual-recovery notice (reached by stopped attempts only now) drops its resolved-cause wording.

• A failed artifact restore during rollback now surfaces instead of silently dropping the
  correction. When safe_rollback restores the preserved BMAD folders from its pre-reset snapshot, a
  genuine git checkout failure (corrupt snapshot, lock, IO) was swallowed alongside the benign
  empty-dir "pathspec did not match" case — so a corrected spec could vanish with no error and loop the
  re-drive. Real failures now raise; the empty-dir case stays tolerated.