v0.7.0 — Fable 5 demo, telemetry, slab stability

Added

• Per-request telemetry: stop_reason + safety-flag logging. Every proxied
  request now records how it ended, so refusal/classifier trips are measurable
  instead of anecdotal.
• Headless bench: multi-turn claude -p driver + events.jsonl scorer for
  fast, non-interactive A/B runs; plus a constant-cost render-style eval harness.
• PXPIPE_DUMP_DIR persists rendered PNGs per request for demo/debug
  inspection of exactly what the model saw.
• Dashboard/factsheet: one-time cache-create losses tagged in the recent
  table; factsheet carries occurrence counts with ticket-style codes.
• Demos: claude-sonnet-5 arm support; fable arm runs claude-fable-5[1m]
  (1M ctx) to match opus/sonnet.

Fixed

• Imaged slab frozen at first render. Volatile content (skill listings, cwd
  caches) stays out of the imaged prefix so turn-2 system sha matches turn-1 —
  no more silent cache-create churn between turns.
• Volatile env text relocated behind all cache breakpoints (not just the
  first), plus cross-session slab stability.
• Refusal-classifier defusing: provenance-framed slab banner and reworded
  tool-docs stub/header — eliminates spurious reasoning_extraction refusals
  on compressed context.
• Render fit to 1568×728 (~1.15 MP) on the Anthropic path for WYSIWYG
  glyphs (what we rasterize is what the model samples).
• Demo cost-ab arms run --no-chrome for reproducible token baselines.

Removed

• compressSchemas knob (superseded by slab stability work).

Docs

• Fable 5 side-by-side demo in the README with verified numbers from the
  recording — same two tasks, same answers: plain $42.21 / 96% context vs
  pxpipe $4.51 — plus the honest caveat (compressed arm needed one nudge for
  single-reply format) and the full attempt log in
  demo/effective-context/ATTEMPTS.md.
• Node transform hook documented as kill-switch only.