feat: ISC map support — the runtime reads the current corpus again - #5

The maps corpus ships only .isc now (zero .imp files remain); the
Python runtime could not parse any of it — every map call raised at
line 1 (CI last ran 2026-08-20, so nobody saw it go dark).

isc.py: a faithful port of the reference TypeScript parser
(interscript-ts src/isc, itself mirroring the Ruby Parslet grammar)
plus the converter into the engine's existing tree — engine and .imp
parser untouched (OCP). Coverage measured against the full corpus:
252/289 maps convert cleanly, zero parse errors; the remaining 37 fail
LOUDLY on known vocabulary gaps (function-in-result 22, etc.) under
the engine's own on_unsupported contract, which now threads through
load_map for conversion-time gaps too.

Converter semantics worth naming:
- parallel subs render to expression source; sequence/bare rules to
  raw regex with before/after/not_before/not_after as lookarounds
  (_guarded_regex) — dropping them was the Oleghovych regression,
  caught by the parity table
- capture-bearing parallel rules degrade to ordered substitutions
  ahead of the capture-free parallel op (documented; the digraph-
  prevention shape every real map uses)
- library dependencies (posix/unicode .iml) contribute aliases,
  harvested and re-rendered through the expression layer — no
  duplicated grammar; single-quoted .iml literals normalized

Parity: bgnpcgn-ukr returns Anton Olehovych / Solomiia / Kyiv and
bgnpcgn-deu returns Tschueß! — byte-identical with the TypeScript and
Ruby runtimes (tests/test_gallery_parity.py is the gate). Bulgarian
98/98 and Greek >=200/242 embedded-test parity restored.