[marcocaggioni]

Skill restructure: Python-first flow-curve-analysis

Restructures the flow-curve-analysis skill so the agent executes via the
Python API (holding result objects as values) instead of shelling out to the
CLI. Same workflow, same judgment content — new hands.

What changed

• Python API is now the primary execution route. Steps 1–5 lead with
  rheofit.print_steps / discover_steps / analyze / fit / plot.
  The CLI stays as the quotable plan artifact (every proposed plan still shows
  the equivalent command) and the one-shot terminal route — now a compact
  CLI reference instead of the main text.
• New "Result objects" contract section, verified against
  analysis.py / models/_fitcore.py: the exact fit() dict keys
  (success, params, redchi, cond, notes, x/y_data/y_fit,
  parent/parent_redchi), the Analysis dataclass fields (results,
  summary columns, outputs, equation, output_dir), and df.attrs
  — plus composition patterns (head-to-head model comparison, seed
  robustness, cross-model summary tables) that only work with values.
• New executable escalation policy (Step 4): when the approved plan names
  ladder-ordered candidates, the agent climbs programmatically —
  redchi > 0.1 + visible residual structure + all params identified →
  next rung; never switches families; nesting violations get one
  thorough/seed+1 retry. No more "offer to climb" round-trips.
• "How to Run" collapsed to importability (uv sync / pip install -e . /
  two-line sys.path fallback). The four CLI fallback routes and the PEP 723
  last-resort are gone from the docs; scripts/flow_curve_fit.py stays on
  disk, unreferenced.
• Step 5 reporting now reads from a.summary / res["notes"] directly;
  Office-workflow tip notes the native-table transcription is direct from values.

Untouched

Fitting Contract, Visualization Contract, model tables (Branch A/B), fixed
Carreau exponents, interview-first golden rule, demo walkthrough (both mermaid
diagrams), material-behaviour reporting guidance, Library Layout, Data Loading
Notes, contributor guides, When to Use, environment packages. Both mirror
copies (.github/skills/ and rheofit/skills/) updated identically.

Checks

• New result-object claims verified against rheofit/analysis.py,
  rheofit/models/_fitcore.py, rheofit/report.py (exact keys/columns)
• All 8 executable Python examples ast.parse clean (3 remaining "failures"
  are the pre-existing intentional sketches, unchanged)
• Fences balanced; no sections dropped (diff reviewed hunk by hunk)

Left unmerged for review.