feat: add monotonic progress bias and pluggable alignment strategy to DemoLibrary (#176)

Add three major improvements to demo-guided alignment:

1. Pluggable AlignmentStrategy protocol with PHashAlignmentStrategy (default),
   CLIPAlignmentStrategy (optional, requires open-clip-torch), and
   HybridAlignmentStrategy (pHash top-K + CLIP re-rank).

2. Monotonic progress bias that penalizes backward jumps in step matching
   to prevent oscillating alignment. Configurable via backward_penalty
   parameter (default 0.3).

3. Adaptive guidance disabling that turns off demo guidance after N
   consecutive low-confidence alignments (default 3 at threshold 0.3),
   preventing the "demo hurts" failure mode.

Also adds AlignmentTraceEntry dataclass for post-hoc analysis of alignment
quality, stored in DemoGuidance.metadata["alignment_trace"].

DemoGuidedAgent now calls reset_alignment_state() on task change and
episode reset to clear monotonic progress tracking.

All 77 tests pass (20 existing + 31 enrichment + 26 new).

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: abrichr