feat(phase-13): natural-language prompt parser + 'ropeway ask' CLI

[harsh-pandhe]

Summary

Phase 13 ships an LLM-style natural-language interface to the optimizer without pinning a third-party SDK or an API key into the test path.

Design choice: deterministic rule-based parser is the primary, fully-tested path. Anthropic SDK / LLM fallback is left as an optional hook gated on ANTHROPIC_API_KEY; CI never exercises it. The CLI surface mirrors this — ropeway ask "build a 2 km MGD with seed 7" works offline.

What the parser recognises (src/ropeway/nlp.py)

• System: mgd, jigback, bgd, 3s, funitel, chair + English aliases (urban gondola → MGD, aerial tram → jigback, tri-cable → 3S, six-pack → chairlift)
• Corridor length: 2 km, 2.5km, 2000 m, two kilometres
• Coordinates: two (lon,lat) pairs separated by comma or space, range-validated against WGS84
• Generations / seed / max towers: num-before-keyword preferred so 80 generations, seed 7 doesn't bleed the seed into the gens field

API

• parse_prompt(text) -> ParsedRequest
• format_summary(req) -> str — human-readable single-paragraph summary
• run_from_prompt(text) -> dict — parsed request + metrics + feasibility
• llm_available() -> bool — True iff ANTHROPIC_API_KEY is set (for the optional fallback)

CLI

ropeway ask "build a 2 km MGD with seed 7"
# Parsed request:
#   system: mgd | corridor: 2000 m synthetic | seed: 7
# Running optimizer on synthetic terrain...
# Feasible: True
#   intermediate_towers: 4
#   cable_length_m: 2161.0
#   ...

Test plan

• tests/test_nlp.py — 27 new tests covering length parsing (numeric + word forms), system aliasing, coordinate pairs, num-vs-keyword precedence, summary formatting, run_from_prompt orchestration
• Full suite 157 → 184 passing, zero regressions
• CLI smoke-tested: ropeway ask parses and runs end-to-end