Skillscript

Closes Perry's f6b479f5 #2 ask. Existing 1924-test suite is heavy on
parse/compile/lint coverage but light on execute-and-assert-output for
the v0.19.12 frozen-surface primitives — without execution-test backing,
the v1.0 surface freeze holds in docs but not in code.

This release adds 58 execute-and-assert tests across five new lanes:

• Lane (a) — tests/v1.0-runtime-semantics.test.ts (26 tests): all
  eight filters in KNOWN_FILTERS — |length (array vs string vs
  object), |trim, |json, |fallback: empty-aware predicate,
  |contains:, |url, |shell, |isodate.
• Lane (b) — tests/v1.0-numeric-coercion.test.ts (9 tests):
  Number()-coerced </>/<=/>= comparisons (not lexicographic),
  if/elif/else chain dispatch, TypeMismatchError surface.
• Lane (c) — tests/v1.0-event-trigger-e2e.test.ts (6 tests):
  Scheduler.fireEvent happy path, EventNotFoundError, strict v1
  param validation (missing + extra), case-insensitive lookup,
  per-fire run_id distinctness.
• Lane (d) — tests/v1.0-composition-runtime.test.ts (5 tests):
  $ execute_skill end-to-end, # Returns: filter, ${R.final_vars.X}
  reach, recursion-depth guard (default limit 10), forward-reference
  resolution.
• Lane (e) — tests/v1.0-output-kinds.test.ts (12 tests): all five
  output kinds (text/none/agent/template/file), Level-2 ${VAR} agent
  target substitution, body-text-as-output template (Pin 4) with var
  interpolation + template-vs-emit channel split, three shell call
  shapes (command= / argv=[...] / unsafe=true), (fallback:)
  op-trailer fire on empty-stdout + no-fire on success.

Single-command invocation: pnpm vitest run v1.0 runs all five lanes
in 200ms. Total: 58 tests, 100% pass on first run after the standard
fixture iteration — no runtime changes needed for the freeze to hold.